All systems

Agent framework built on OpenAI Assistants API — agencies (roles + comm flow); replaced after Assistants v2 changes.

Google-led open protocol for agent-to-agent collaboration (April 2025) — agent cards, RPCs, server-sent events. Partner companies: Atlassian, MongoDB, Salesforce, ServiceNow, etc.

Cisco + LangChain-led 'Internet of Agents' collective — open standards for discovery, messaging, observability of agent networks.

Claude 3.5 Sonnet's beta capability for controlling a computer (mouse, keyboard, screen). Frame for browser-agent and OS-agent products built on Claude.

Authentication and tool platform for AI agents — fully-managed OAuth flows for LLMs to act on user accounts (Gmail, Slack, etc.).

Originally Microsoft Research's multi-agent framework — fork branded AG2 (2024-10) led by community. Memory is one optional component (AutoGen Memory cross-listed).

The first viral autonomous agent demo (March 2023) — GPT-4 + memory + tool use + self-direction. Significant Gravitas raised $12M to commercialise the AutoGPT Platform.

Yohei Nakajima's iconic spring-2023 task-list agent loop — Pinecone-backed memory, OpenAI for execution; ~150 LOC. Influence-disproportionate-to-code.

DAGWorks's state-machine framework for agents — state + actions + transitions; lightweight alternative to LangGraph.

Open-source autonomous coding agent for VS Code — terminal + browser + edit operations; uses Anthropic Computer Use style flow.

Authentication and tool-execution platform for AI agents — 100+ apps with OAuth handled, function-call schemas, observability.

Open-source AI coding assistant — VS Code / JetBrains; configure-your-own-models. Existing Continue-rules row covers file-backed memory; this row covers the framework.

Multi-agent framework — Crews (sequential) + Flows (event-driven). Open-source core + CrewAI Enterprise. Memory available via Mem0 / Letta adapters (CrewAI Memory cross-listed).

ai16z's TypeScript agent framework popularised in the crypto/Web3 agent space — multi-platform (Discord, X, Telegram). Brand evolved into elizaOS.

One of the original 'agent writes a whole codebase from a prompt' demos — community-maintained; spawned the commercial Lovable (gptengineer.app).

deepset's open-source framework for LLM/RAG/agent apps — Pipelines abstraction; deepset Cloud commercial. Memory is a downstream connector.

Event-driven durable workflow platform — competes with Temporal for the JS/TS/Python crowd. Inngest AgentKit (memory cross-listed) is the agent extension.

Dominant Python/JS framework for LLM apps — tool calling, agents, retrievers, runnables. Memory is one component (LangMem cross-listed); the framework itself is the product.

LangChain's stateful graph runtime for agents — explicit state, persistence, time-travel debugging, human-in-the-loop. Memory is one persistence target (LangGraph Persistence cross-listed).

Data framework for LLM apps — connectors, indices, query engines, agents, workflows. Memory is one of many subsystems (LlamaIndex Memory cross-listed).

Microsoft Research's generalist multi-agent system — Orchestrator + WebSurfer/FileSurfer/Coder/ComputerTerminal specialists. Built on AutoGen v0.4.

Prefect's LLM toolkit — turns Python functions into LLM-powered ones; structured outputs, classification, extraction. Replaced by Control Flow for agentic workflows in 2024.

TypeScript-first agent framework — Workflows, Agents, RAG, eval. Built by the Gatsby team. YC W25.

Anthropic-led collection of reference Model Context Protocol servers (filesystem, GitHub, Slack, etc.) — substrate beneath the entire MCP-memory ecosystem (Cipher, Memobase, etc.).

Multi-agent framework simulating a software company (PM, Architect, Engineer, QA). Famous for its software-development-via-multi-agent demo.

Microsoft's enterprise-focused agent framework — multi-language (C#, Python, Java); Plugins / Planners / Memory; integrates with Azure AI / Foundry.

Anthropic-introduced open protocol (Nov 2024) for connecting LLMs to data sources and tools. Now standard across Anthropic, OpenAI, Microsoft, Cursor, Continue, IDEs, etc.

Open-source workflow automation platform with first-class AI Agent / LangChain / Memory / Tool nodes — fastest-growing workflow tool in the AI era.

OpenAI's official Python framework for agents — handoffs, guardrails, tracing built in. Successor to OpenAI Swarm (experimental); officially supported.

OpenAI Swarm — experimental educational multi-agent framework released Oct 2024; deprecated in favour of the OpenAI Agents SDK in Mar 2025.

Open-source SWE agent platform (renamed from OpenDevin to OpenHands). All Hands AI raised $5M; widely used as the OSS counterpart to Devin.

Multi-modal agent framework (renamed from Phidata to Agno in 2025) — agents, memory, knowledge, tools; emphasis on multi-modal pipelines.

Pydantic team's agent framework — type-safe agent definitions, structured outputs, tool calls validated by Pydantic schemas.

Durable-execution platform from ex-Confluent / Flink team — Restate Cloud + OSS server; emphasises low-latency invocation vs Temporal's polling.

Hugging Face's minimal agent library — emphasises code-execution agents (agent writes Python, runs it) over JSON tool-calling. Library is intentionally ~1k LOC.

AWS-released open-source agent framework — released March 2025; model-driven, simple Agent abstraction. Predecessor of AWS Agent SDK / Bedrock AgentCore.

Durable-execution platform — Workflows-as-code with crash safety. Used as agent-orchestration backbone at Stripe, Snap, Coinbase, OpenAI Stargate.

Open-source JS/TS background-jobs platform — long-running, retries, observability. Trigger.dev v3 (2024) added 'tasks' abstraction for agent workflows.

Vercel's TypeScript LLM SDK — streaming, tool calls, structured outputs, agents. Dominant framework for Next.js LLM apps.

OSS terminal coding agent by Paul Gauthier — one of the original AI-pair-programmer CLIs (2023). Architect/Editor modes; repo-map for context; uses any LiteLLM-compatible model. Memory via CONVENTI…

Amazon's developer-focused agent — formerly CodeWhisperer (renamed 2024-Q4). Available as IDE plugins (VS Code, JetBrains, Visual Studio, Eclipse), in AWS Console, and CLI. Strong AWS-tooling integ…

Anthropic's API-surface capability that lets Claude control a virtual desktop — see screenshots, type, click. Launched 2024-10 as a public beta. Not a finished IDE itself; it's the substrate that o…

AWS Bedrock AgentCore — managed runtime to deploy + scale + observe agents. Multi-component: Runtime, Identity, Memory, Gateway, Browser, Code Interpreter, Observability. Distinct from Bedrock Agen…

Cross-listing of Bolt.new at the harness layer. StackBlitz's WebContainer technology runs full Node.js in-browser; Bolt prompts the user for an app spec, then writes/edits/runs/deploys it entirely …

Charm Bracelet's terminal-native AI coding assistant — built in Go atop the Bubble Tea TUI framework. Multi-model via Glamour/Mods integration. Open-sourced 2025; positioned as a TUI-native alterna…

OpenAI's cloud-hosted, asynchronous coding agent surfaced inside ChatGPT — distinct from Codex CLI (terminal) and from the deprecated 2021 Codex API. Released 2025-05. Reads + writes from sandboxed…

Anthropic's official agentic coding CLI — runs Claude as a long-running coding agent in the developer's terminal. Memory layer = CLAUDE.md (project + user scoped), agent skills, MCP servers. Releas…

OSS VS Code extension that runs as a fully agentic coding assistant — plan/act modes, file ops, shell exec, browser-use, MCP. Originally "Claude Dev" (2024); rebranded Cline 2024-Q4. Multi-model: A…

Cross-listing of Cognition's Devin product framed at the harness level rather than the memory layer. Devin 2.0 (Mar 2025) added cheaper concurrent sessions; Spec Mode (Aug 2025, post-Windsurf-IP ac…

OSS in-IDE coding-assistant plugin — supports VS Code and JetBrains. Customisable: any model, any context provider, any slash-command. "Agent mode" added 2024 for multi-step plan + execute. Memory …

AI-first IDE that forked VS Code; Composer (multi-file edit) and Agent Mode (long-running coding agent) are its flagship surfaces. Memory layer = .cursorrules / Rules; recall via repo-aware Context…

The original AI coding assistant (Copilot in editor; 2021). 2025 evolved into a full agentic surface: chat, completions, Agent Mode (multi-step), and Copilot Workspace. Backed by Microsoft. Distinc…

GitHub's hosted agentic workspace — given an issue, generates a spec, plan, and PR. Preview launched April 2024; rebranded "GitHub Copilot Coding Agent" + "Copilot Agent Mode" in 2025 with broader …

Block (Square)'s open-source on-machine agent — terminal + desktop GUI. Modular: extensions for shell, computer-control, MCP servers; multi-model via LiteLLM. Released 2024; positioned as a general…

JetBrains' in-IDE assistant integrated across IntelliJ, PyCharm, GoLand, WebStorm, Rider, etc. Multi-model: OpenAI, Anthropic, Google, local. Differentiated by its deep IDE-API integration — refact…

JetBrains' answer to Cursor Agent Mode / Devin — an autonomous in-IDE coding agent. Public beta 2024-Q4; runs alongside AI Assistant inside IntelliJ-platform IDEs. Plans, edits, runs tests; multi-m…

Spec-driven agentic IDE from AWS — preview launch July 2025. Codifies the AI-coding workflow as Requirements / Design / Tasks files generated before code ("specs"); also persists "steering" files a…

Cross-listing of Lovable at the harness layer. Swedish startup (Stockholm) that grew out of GPT Engineer (Anton Osika). Builds full-stack React/Supabase apps from prompts; AI handles git, deploys, …

AI-coding startup focused on extremely long-context models (claimed 100M tokens) and a frontier coding agent. Raised $320M Series B from Eric Schmidt, CapitalG, Atlassian Aug 2024. Lower public-pro…

Chinese-origin general-purpose autonomous agent launched March 2025 by Butterfly Effect. Operates in a cloud sandbox — browser, file system, shell. Went viral on launch with invite-only access; pos…

Cross-listing of MetaGPT as a harness. Multi-agent OSS framework that simulates a software company — PM, architect, engineer, QA. Generates docs + code + tests from a one-line prompt. ~44k★. Distin…

OpenAI's open-source terminal coding agent — released April 2025 alongside o3/o4-mini. Counterpart to Claude Code in the OpenAI stack. Memory via ~/.codex/memories (file-backed); routes to OpenAI o…

SST team's open-source terminal coding agent — TypeScript, multi-model, MCP-native. Released 2025; explicitly positioned as a permissively-licensed Claude Code alternative. Memory via AGENTS.md (th…

Cross-listing of OpenHands at the harness layer. OSS Devin-alternative — sandboxed full SWE agent with browser + shell + editor. Renamed from OpenDevin Mar 2024. ~40k★. Commercial offering: All Han…

Cross-listing of Replit Agent at the harness layer. Replit Agent v3 (2025) added background agents + agent-driven deploy. The Coding-agent-memory row remains canonical for the memory framing.

OSS VS Code agentic extension — Cline fork that diverged with custom modes (Architect, Code, Ask, Debug). Memory via .roorules + Memory Bank. Smaller community than Cline but persistent contributor…

Early (2023) OSS coding agent by Swyx — minimalist single-prompt-to-codebase loop. ~12k★. More artifact-than-product now (subsumed by Cline / Aider / Claude Code) but historically significant — def…

Sourcegraph's coding assistant — IDE plugin (VS Code, JetBrains) plus web. Distinguishes itself with Sourcegraph's code-search index as the context provider (cross-repo, codebase-wide retrieval). A…

Vercel's AI product (formerly v0.dev) that generates React/Next.js components and full apps from natural-language prompts. Tight integration with Vercel deploy; uses Vercel-hosted models (multi-LLM…

Codeium-built agentic IDE (VS Code fork); flagship was Cascade — a planning + execution agent that maintains "Flows" and "Memories" as durable artifacts. Reported acquired by OpenAI in May 2025 for…

Native (Rust) collaborative code editor by Zed Industries (founded by the Atom + Tree-sitter team). Added Agentic Editing in 2024 — multi-step coding agent driven by Claude or local models; .rules …

Anchor Browser provides cloud-hosted Chromium instances optimised for AI agent use — anti-detect, persistent sessions, residential proxies, CAPTCHA handling. Founded 2023, YC W24. $2.6M seed disclo…

Browserless is a headless-browser-as-a-service — managed Chromium / Playwright / Puppeteer endpoints for web scraping, screenshots, PDF generation, web automation. Increasingly used as substrate fo…

Coder.com (formerly Coder Technologies) provides enterprise self-hosted cloud development environments — VS Code Server + dev workspaces on K8s / VMs. $30M Series B 2021 (Redpoint + Founders Fund +…

CodeSandbox is the long-running browser-based dev sandbox (founded 2017 by Ives van Hoorne + Bas Buursma in Amsterdam). **Acquired by Together AI Oct-2024** to anchor Together's agent-execution sta…

Daytona is an open-source cloud development environment (CDE) platform — self-hostable alternative to Gitpod / GitHub Codespaces. 2024 added AI-agent sandbox tier for ephemeral agent code execution…

e2b provides secure, ephemeral sandboxes (Firecracker microVMs) for AI agent code execution. Each sandbox spins up in <200ms with full Linux environment, Python/Node/etc preinstalled. Substrate for…

Gitpod is the original cloud development environment (CDE) — founded 2014 by Sven Efftinge + Johannes Landgraf (also TypeFox / Eclipse Theia / Open VSX). $25M Series A 2022 led by General Catalyst.…

Modal Labs ships a serverless container platform with first-class GPU support; Sandboxes feature (2024) adds ephemeral, AI-agent-friendly compute sandboxes. Founded 2021 by Erik Bernhardsson (ex-Sp…

StackBlitz's WebContainers technology — browser-native Node.js runtime running fully client-side via WASM. Powers Bolt.new (already in catalog) plus many browser-based AI dev experiences. Founded 2…

Workato is the enterprise iPaaS (integration platform) leader — $200M+ ARR, $5.7B valuation 2021. 2023-24 launched Workato AI / Workbot AI with AI-agent workflow sandbox capability inside its autom…

Privacy-focused integrated assistant. Memory is user-curated: explicit facts and preferences persist across all sessions, stored locally with no server retention. Agentic browsing experiment (Brave…

Cloud browser infrastructure for AI agents. Headless Chromium sessions with persistent cookies, localStorage, session recording — agent workflows carry state across runs. Memory at the session/auto…

The Browser Company's AI-native browser; acquired by Atlassian for $610M (Sept 2025). Stores summaries and preferences locally; opt-in @history feature references up to 7 days of browsing as contex…

Silicon Valley. Spatial agentic browser with explicit Agentic Memory layer learning from history and notes for instant recall of past research sessions. Operates on logged-in accounts (Quora, X, et…

Copilot Mode (Oct 2025): Journeys groups related browsing across sessions into resumable project cards with AI summaries. Copilot Actions execute multi-step tasks inside the browser. Assistant can …

Memory feature stores user-declared preferences and facts; injected into future prompts. Understands current tab content; manages open tabs contextually. Memory off by default, end-user controlled,…

AI-native Chromium-based browser. Assistant operates across tabs and sessions; persistent memory of preferences; multi-step agentic tasks without leaving the browser. Memory is opt-in and server-si…

Standalone browser with fully on-device LLM. Agent acts inside websites (logging in, clicking, completing tasks); memory-based actions recall past searches to auto-fill similar tasks with no cloud …

Background sub-agent in Claude Code that periodically scans ~/.claude/projects/<project>/memory/ + recent JSONL transcripts via grep, prunes stale notes, deduplicates, resolves contradictions, conv…

Persistent file-based memory at ~/.claude/projects/<project>/memory/ with a MEMORY.md index plus typed memory files (user / feedback / project / reference). Loaded into context across sessions.

Claude Code plugin packaging of Memvid (see Dedicated memory layers). Wraps the Memvid .mv2 single-file store as a slash-command-installable plugin so per-project memory becomes a git-committable f…

FastAPI server. RBAC + API-key auth, encryption, backup/recovery, embedding pipeline with deduplication. Only server in this list with role-based access control over pgvector.

Hooks into Claude Code's session lifecycle; compresses tool outputs into ~500-token observations via Claude's agent SDK. SQLite + FTS5 with optional Chroma vector. "Endless Mode" compresses in-flig…

Wraps Supermemory cloud API. Team memory shared across project members, separate from personal memories. Auto-capture on session end; per-repo config.

High-performance MCP server building persistent code knowledge graphs via AST. 155 languages via vendored tree-sitter; SQLite + WAL; single static binary; zero runtime deps. Knowledge graph: Projec…

"Cognify" extracts entities and relationships into a structured graph. Saves interaction patterns as coding rules; uses both vector and graph DBs. Bundled in Cognee v0.3.5+.

SQLite + FTS5; agent decides what to save (not firehose capture). Single Go binary, no CGO, no Node/Python/Docker dependencies. HTTP API + MCP + CLI + TUI; optional cloud sync with local-source-of-…

Tracks how facts change over time; maintains provenance; supports prescribed and learned ontology. Bi-temporal edges (when known vs when valid). Only server that distinguishes "fact learned" from "…

World (facts), Experiences (agent actions), Opinion (beliefs with confidence scores), Observation (mental models). Entity resolution ("Alice" and "my coworker Alice" same node). Mimics human recall…

SQLite + FTS5 + BM25 ranking. Stores learnings, decisions with reasoning and alternatives , entities, observations. Loads context from last 3 sessions. Only server that captures why a decision was …

KG-based local fork of @modelcontextprotocol/server-memory . Avoids ephemeral npx installs; supports named databases for topic scoping; user-controlled memory path. Drop-in for teams that want the …

Embeddings via @cf/baai/bge-m3 ; Vectorize + D1; per-user namespace isolation. Runs entirely on Cloudflare Workers — no separate server.

SQLite-Vec default, Cloudflare hybrid for production. Typed edges (causes, fixes, contradicts). Autonomous consolidation compresses old memories. SSE real-time notifications + D3.js visualisation d…

Mem0's managed cloud API exposed as MCP tools. Semantic search + CRUD on memories scoped by user / agent / session. Cloud-hosted MCP endpoint — no local infrastructure required.

Knowledge-graph-based persistent memory MCP server — the canonical reference implementation. npx -y @modelcontextprotocol/server-memory ; configurable via MEMORY_FILE_PATH .

Local-first self-hosted Mem0 variant. Shared memory layer bridges Claude Desktop, Cursor, and Windsurf in one session. All data stays on-machine. Docker Compose self-host.

Claude Code plugin that indexes prior conversations at session end and exposes search/show tools via MCP. Distributed via obra/superpowers-marketplace ; install with /plugin install episodic-memory…

Appends conversation turns to .aider.chat.history.md in project root. restore_chat_history defaults to false — no automatic replay into new sessions. Community /session save command for JSON state …

Real-time knowledge graph indexed from connected repos + commit history + CI/CD + tickets + docs. Separate Agent Memory layer (Sept 2025) captures session decisions/corrections/goals as discrete it…

In-browser WebContainer (Rust-backed shared memory) is the implicit state store. No dedicated memory abstraction — project files are the state. Team plan adds shared workspaces + GitHub integration.

Official Memory MCP (Mar 2025) backed by a knowledge graph. Memory is injected via MCP rather than native to the Continue core. Scope depends on MCP server deployment (local or shared).

Structured Knowledge items stored server-side. Three-tier scope: repo-pinned (always injected), global (recall by relevance), org/enterprise (visible to org members). Auto-generates Repo Knowledge …

Long-term context stored as files in /docs that the agent reads on each session. Agent Mode (default since July 2025) holds working context within session but doesn't persist cross-session beyond f…

Vibe is Mistral's cloud-hosted coding-agent platform: each session runs in an isolated sandbox with task state, history, and approvals. Spawnable from CLI or Le Chat, with parallel execution. Local…

Memory under ~/.codex/memories/ : summaries, durable entries, recent-input logs, supporting evidence — generated in background after sessions go idle. Chronicle (opt-in macOS preview) uses screen-c…

No documented built-in persistent semantic memory store. Session duration expanded to 200 min (Agent 3); Agent 4 shifts toward persistent runtimes. Temporal used for orchestration durability, not s…

No dedicated persistent memory layer. Operates as a GitHub App — reads issue + codebase fresh on each trigger. Some context implicit in the PR/issue threads it created previously.

Enterprise Context Engine builds a multi-dimensional knowledge graph from repos + CI/CD + PRs + tickets + docs at the org level. Individual completions stay session-scoped. Tabnine Agents (Nov 2025…

Python library that turns any LLM into a browser agent — extracts DOM into a structured representation the LLM can act on (rather than pure screenshot). Browser Use Inc. raised $17M seed Q1 2025 (F…

Open-source desktop-agent runtime that runs an LLM-controlled Linux desktop inside Docker — uses xdotool / X11 capture for control. Pitched as the 'self-hosted Operator' alternative; agent sees a r…

Anthropic's Chrome extension that lets Claude act inside the user's existing browser (rather than a sandboxed one). Research preview launched August 2025 — limited to Claude Max tier. Distinct from…

Always-on desktop AI assistant that watches the user's screen (across all apps) and answers questions / takes actions based on what's currently visible. Mac + Windows. Founded by Mac Bishop (ex-Mix…

Cloud-hosted headless browser API tuned for AI agents — captcha solving, stealth mode, session recordings. Y Combinator S24; pitched as a faster / cheaper Browserbase alternative. Used as substrate…

AI agent platform that automates SaaS workflows across Gmail, Slack, Sheets, etc. — uses natural-language instructions that compile to durable 'playbooks.' Founded by ex-Google/Facebook engineers; …

Open-source browser-agent framework focused on AI-driven end-to-end testing — natural-language test cases that translate to Playwright actions. Y Combinator 2025; positions itself as 'Playwright + …

OpenAI's general-purpose browser agent — research preview launched January 2025 (US Pro tier). Built on the Computer-Using Agent (CUA) model that processes screenshots and emits mouse / keyboard ac…

Google DeepMind's experimental browser agent built on Gemini 2.0 — Chrome extension that takes actions on the user's behalf. Research-preview launched Dec 2024. The Google answer to OpenAI Operator…

Open-source autonomous browser agent that uses computer vision + LLMs to navigate websites. Workflow-based: agents follow recorded plans that adapt to layout changes. Skyvern Inc. raised $4.5M seed…

Open-source TypeScript framework for browser automation, maintained by Browserbase. Hybrid of Playwright-style imperative APIs and AI-driven natural-language actions (e.g. page.act('click the sign-…

Open-source browser API ('the browser for agents') — Chromium fleet with stealth mode, session recordings, CDP. OSS repo + managed cloud. Compatible with Playwright / Puppeteer. Distinguishes itsel…

ByteDance Seed's open-source GUI-agent foundation model — released Jan 2025; trained end-to-end on screenshot → action pairs. Comes with UI-TARS-Desktop (Electron app) that runs the model against t…

SEA-LION-Embedding (March 2026): retrieval + reranking models contrastively trained on 245M text pairs across 10 SE Asian languages. SEA-BED benchmark (169 datasets). SEA-LION v4 (Gemma-based) at 1…

German enterprise stack. PhariaCatch captures + preserves organisational knowledge through structured human-machine collaboration; PhariaAssistant queries via semantic search. Combines sovereign de…

Native graph-based memory for autonomous AI agents. Bio-inspired sleep-based memory consolidation.

Tsinghua spinout. Foundation agent series for autonomous smartphone/PC GUI control. Persistent task memory retains shopping history, app usage patterns, multi-step task state across sessions. AutoG…

Agent memory product evaluated against LoCoMo.

Agent-memory product. LLM-curated hierarchical context.

Korean startup. Connects scattered internal/external data across multiple chats and services to improve AI answer quality at retrieval. 18-year-old CEO Jang Jun-ha; pitch is that answer quality dep…

Opensource memory layer focused on programming agents — works with Cursor, Windsurf, Claude Code, Roo Code via MCP. Maintains a coding-context memory across sessions and across IDEs.

"Extract–Cognify–Load" pipeline that turns raw input into a typed, queryable knowledge graph for agent recall.

Standalone memory service from Vectorize. Open source. Biomimetic four-network design (World, Bank, Observation, Opinion). Ships an MCP memory server.

Connects Slack, Gmail, Notion, Drive so agents can recall and reason across company knowledge.

German startup. Captures how real work is performed across enterprise systems and teams; converts process knowledge into a persistent memory layer agents use to automate complex workflows. Targets …

Israeli production model. Interleaves Mamba SSM layers with Transformer attention (1:7 ratio) + MoE blocks every 2 layers. SSM layers maintain fixed-size recurrent hidden state — compressed long-co…

India's first agentic AI assistant. Persistent memory of prior interactions across sessions; accepts voice/text/image/document input building continuous context. Decomposes goals into sub-tasks exe…

LLM-as-Operating-System paradigm — model manages its own memory, paging between core context, recall buffer, and archival store. UC Berkeley origin; transitioning to letta_v1_agent .

ByteDance Seed + Zhejiang Univ + SJTU. Multimodal agent with long-term memory as entity-centric graph built from live video + audio streams. Parallel memorisation + control processes. Ships M3-Benc…

Universal memory layer for AI agents. Three concurrent stores (vector + graph + KV); LLM-extracted facts; concurrent retrieval via ThreadPoolExecutor.

Open-source long-term memory framework for autonomous agents.

Multi-layered open-source memory system for AI agents.

User-profile-based long-term memory. Per-user profile + event timeline. API + Python/Node/Go SDKs + MCP server. Profile-centric rather than fact-centric.

Shenzhen. Argues attaching external vector stores is architecturally flawed; trains memory as an intrinsic capability via Agentic RL. Plans B2B enterprise KM and C2C personalisation in 12 months. F…

Structured, persistent memory layer for AI agents. Open source.

SQL-native memory engine for AI agents — single library, opt-in conscious/auto modes, structured Pydantic outputs. Aims to be 'one line of code' to add memory to any LLM workflow.

Memory layer from MatrixOrigin. Strong-retrieval / clear-reader-separation design.

Memory layer for AI assistants — semantic short-term and long-term memory with concept-graph linking and decay/reinforcement.

YC P26 (Spring 2026). Unified memory layer so all agents on a team share one persistent context — meetings, Claude sessions, Slack threads all become queryable. Solves the "goldfish brain" problem …

Single-file memory layer (one .mv2 file). No DB, no server. Append-only sequence of immutable Smart Frames with timestamps + checksums. Native Rust core (rewritten from Python).

Walrus is a Sui-blockchain decentralised storage protocol; MemWal is its AI-facing SDK that stores encrypted agent memory on-chain — portable, verifiable, shareable across agent systems with no sin…

Reasoning-focused successor to HyperCLOVA X. Pre-trained on ~6T high-quality Korean + English tokens. 128K context via rotary position embeddings + grouped-query attention. Sovereign reasoning LLM …

Builds agents that learn from experience by constructing a "world model" of the micro-environment they operate in. Specialise over time through Agentic RL loop rather than relying on static prompt …

AI-native storage tier for gigascale agentic inference. Bridges GPU HBM and general-purpose storage with petabyte-scale RDMA-accelerated KV-cache placement; coordinated via DOCA, Spectrum-X Etherne…

Deploys agents across the web to build structured profiles of individuals from public digital footprint, then feeds the people graph into AI agents as contextual memory so they understand users bef…

Omni memory system for personalised, long-horizon, self-evolving agents. From OPPO PersonalAI.

Memory engine product. Pitch is high-accuracy retrieval at very low latency.

Open-source cognitive memory engine for LLMs and agents.

Long-term memory engine for developers — captures everything the user does (browser, IDE, terminal) on-device, exposes via MCP for any IDE/agent.

Cloud-free agent memory using information-geometric retrieval (Fisher-Rao distance + BM25 + entity graph + temporal fusion via 4-channel RRF) instead of LLM calls for scoring and lifecycle. Mode A …

Memory engine with API, app, browser extension, and MCP server. Extracts facts, tracks updates, resolves contradictions, auto-forgets expired info. Plugins for Claude Code, OpenCode, OpenClaw, Hermes.

YC S25 + $3M seed (Feb 2026). Workflow orchestration mapping corporate environments as a structured context graph so agents always have institutional knowledge for tasks. CTO frames differentiator …

Node.js/TypeScript-native memory layer. SQLite + local Transformers.js embeddings (~80MB model via WebAssembly). AUDN loop (Add/Update/Delete/None) resolves contradictions before write rather than …

Bi-temporal knowledge graph (event time + ingestion time). Strong on chronological reasoning and contradiction tracking. Graphiti is the open-source core.

BAAI (Beijing Academy of AI) BGE family — bge-large/base/small + bge-m3 multilingual + bge-reranker. One of the most widely-used open embedding families.

Cohere's hosted embedding models (embed-v3 family) — multi-lingual, image+text, compression. Reranker API (Rerank-v3). Cohere is one of three major LLM startups (alongside Anthropic, OpenAI) by rev…

Alibaba DAMO Academy's GTE family — open-weight multilingual embedding models; gte-large/base/small + Qwen-based gte-Qwen2.

HF's hosted Inference API and Inference Endpoints — serve any model from the Hub including the BGE/GTE/E5 embedding families.

Jina AI's embedding family — jina-embeddings-v3 (multilingual), jina-clip-v2 (text+image), jina-colbert-v2. Open-weights + hosted API + ColBERT-style.

Mistral AI's hosted embedding API — mistral-embed; available via Mistral La Plateforme and Azure AI.

Hosted embedding and reranker APIs — mxbai-embed-large + reranker. Open-weight models + paid API + agentic search.

Nomic's open-weight embedding models (nomic-embed-text-v1.5, nomic-embed-vision). Nomic Atlas for vector dataset exploration.

OpenAI's hosted embedding API — text-embedding-3-small, text-embedding-3-large. Most widely used commercial embedding API in OSS RAG/memory stacks.

Reimers & Gurevych's library for training sentence embeddings — substrate under most open-weight embedding models. Now under Hugging Face.

Specialised embedding and reranker provider — voyage-3, voyage-multilingual, voyage-finance, voyage-law, voyage-code. Acquired by MongoDB Feb 2025 for $220M.

NeuralSearch combines vector + keyword via neural hashing — compresses to 1/10th size while retaining 99% info. AI-powered personalisation + recommendations.

Bibliographic metadata curation (Web of Science, Derwent, Cortellis). Human editorial governance + journal-deindexing. Memory-adjacent — included as a curated-knowledge baseline.

RAG-as-a-Service for AWS (Dec 2025) via hosted MCP server grounding Amazon Bedrock agents in enterprise knowledge. Passage retrieval + answer generation + ranked search + fetch in one API.

Enterprise search with 100+ connectors. Personalised per-user knowledge graph. No governance layer.

Enterprise Q&A agent powered by Luci patent-pending ultra-precise RAG. Embeds on product detail pages; consumes technical PDFs, spec sheets, images, tables, charts, graphs and product manuals. Main…

Semantic + hybrid search GA (2025). Automatic embedding generation + caching via OpenAI / HuggingFace / Ollama. Multi-modal (text + images); hybrid rank fusion; conversational RAG built in.

Hybrid keyword + vector retrieval with entitlement-aware filtering. Unified enterprise knowledge graph linking documents, tickets, records. RAG prompt orchestration built in.

Agentic enterprise search with Knowbler — quality checks via the KCS (Knowledge-Centered Service) standard. Most explicit governance features in the segment.

Hybrid retrieval combining dense vector + keyword + graph traversal + multimodal. Agentic RAG grounding AI agents in internal knowledge. SOC 2 / ISO 27001 compliant.

Production ML monitoring + LLM observability (Arize AX) — drift, eval, root-cause. Pre-LLM observability leader; now LLM-pivoted.

Arize AI's open-source LLM observability tool — OpenTelemetry-native LLM tracing, evals, datasets. Arize commercial offers Phoenix + Surveyor.

Eval-first LLM platform — datasets, eval runs, prompt playground, tracing. Founded by ex-Impira (acquired by Dropbox) team. Used by Notion, Brex, AirOps.

Open-source DeepEval framework + Confident AI cloud — pytest-style eval runs, hallucination/relevance/bias metrics, datasets.

Galileo Evaluate / Observe / Protect — eval, tracing, guardrails. Big focus on hallucination detection and ChainPoll evaluator metric.

Open-source LLM observability with one-line proxy integration — gateway routes through their proxy, captures every LLM call. Sessions + custom properties + evals.

Collaborative LLM development platform — non-eng prompt editor, eval, tracing. UK; Index Ventures + Y Combinator backed.

Arize-led open OTel semantic conventions for LLM/agent tracing — auto-instrumentation for OpenAI/Anthropic/LangChain/LlamaIndex etc.

Traceloop's open-source OpenTelemetry auto-instrumentation for LLM workflows — works with any OTel-compatible backend (Datadog, New Relic, Grafana, etc.).

Enterprise eval and guardrails platform — Lynx hallucination detector (open weights), Glider judge model, Polyglot benchmark. Built by ex-Meta researchers.

Prompt versioning + tracing platform — non-eng-friendly prompt editor; early entrant pre-LLM-ops wave.

Open-source eval framework for RAG pipelines — faithfulness, answer-relevance, context-precision, context-recall. Widely used as a baseline RAG eval kit.

ML observability with the TruLens open-source LLM eval framework — the 'RAG Triad' (context-relevance, groundedness, answer-relevance) metrics. Acquired by Snowflake 2024-05.

Open-source LLM/RAG evaluation framework — feedback functions, the RAG Triad. Originally TruEra; now Snowflake-stewarded.

End-to-end LLM development platform — eval, prompt management, deployment, observability. Targeted at enterprise teams shipping LLM products.

AGENTS.md at repo root (and ~/.codex/AGENTS.md for global); hierarchical concatenation from root to cwd. Persists build/test commands, conventions, architecture overview, security constraints, git …

CONVENTIONS.md (any name, loaded via --read or read: in config); .aider.conf.yml (home / git-root / cwd, last wins); .aiderignore (gitignore-syntax).

Markdown file injected at session start; reloaded post-compaction. Document-as-memory: human-readable, version-controlled, no extraction or vector store. Anthropic-canonical pattern; equivalent eme…

.clinerules single file or .clinerules/ directory of numbered .md files (e.g. 01-style.md ). Global at ~/.cline/rules/ . Files version-controlled and AI-editable — Cline can write back. Toggleable …

memory-bank/ directory of Markdown: projectBrief.md , productContext.md , activeContext.md , systemPatterns.md , techContext.md , progress.md . Loaded in full at start of every task. Compensates fo…

.continue/rules/*.md at workspace root. YAML frontmatter: name , globs (file-pattern trigger), alwaysApply , description . Loaded lexicographically.

.cursorrules + project rules. Semantic chunking + Merkle-tree incremental re-indexing over the codebase. Turbopuffer-backed retrieval at IDE scale.

.github/copilot-instructions.md repo-wide; .github/instructions/<name>.instructions.md for path-scoped rules with YAML frontmatter ( applyTo glob). Persists style, test conventions, language prefer…

replit.md at project root. Persists project overview, tech stack, architectural decisions, deployment process, external resources. Scoped to Agent conversations only.

.roo/rules/ (recursive .md / .txt alphabetic load); .roorules single-file fallback. Mode-specific via .roo/rules-{modeSlug}/ . Global at ~/.roo/rules/ . Multi-tenant mode customisation.

.vscode/cody.json (workspace-scoped). JSON defining slash-commands with prompt text, context options ( currentFile , openTabs , terminal ), and mode ( ask / edit / insert ). Reusable prompt templat…

Rules files plus Cascade auto-generated memories. "Fast Context" RAG using SWE-grep models; ~100K-token effective context.

.rules at project root. Falls back to reading .cursorrules , CLAUDE.md , AGENTS.md (first match wins). Plain Markdown, no frontmatter schema. Injected as system prompt for all Agent Panel interacti…

Kai-Fu Lee's 01.AI (founded 2023, Beijing). Yi family open-weights — Yi-34B (2023), Yi-Large (proprietary 2024), Yi-Lightning (2024-10, frontier-tier, matched GPT-4o on LMSYS chatbot arena). Apache…

Alibaba Cloud's Qwen 3 family (2025-04) — open-weights frontier models from 0.5B to 235B-A22B MoE. Apache 2.0 (most sizes) — broadest open-weights family. Includes Qwen 3 32B dense, Qwen 3 30B-A3B …

Amazon's first-party foundation model family, launched at AWS re:Invent Dec-2024. Nova Pro (multimodal mid-tier), Nova Lite (multimodal cheap), Nova Micro (text-only fast), Nova Canvas (image gen),…

Claude family of frontier foundation models — Sonnet 4 (workhorse), Opus 4.5 (frontier capability), Haiku 4.5 (fast/cheap). Frontier model used as substrate by Claude Code, Anthropic Memory tool, C…

Toronto-based foundation-model lab (Aidan Gomez, ex-Google Brain 'Attention Is All You Need' co-author). Command family is enterprise / RAG-optimised: Command A (Mar-2025, 111B dense, 256k ctx), Co…

DeepSeek (Hangzhou-based, Liang Wenfeng founder) — released DeepSeek-V3 671B MoE (37B active) Dec-2024, then R1 reasoning model Jan-2025 that matched o1 at ~3% of inference cost. Caused 'DeepSeek m…

Google DeepMind's Gemini 3 frontier model family — Gemini 3 Pro (top tier), 3 Flash (fast), 3 Nano (on-device). Substrate for Google Workspace AI, Vertex AI Agent Platform, Gemini Code Assist, Proj…

Meta's Llama 4 family (released 2025-04) — Scout (109B MoE, 17B active), Maverick (400B MoE, 17B active), Behemoth (2T MoE, 288B active; preview). First Llama generation to use mixture-of-experts a…

Microsoft Research's Phi family of small, data-curation-focused open-weights models. Phi-4 (14B, Dec-2024) matches much larger models on STEM benchmarks. Phi-4 Multimodal (5.6B, Feb-2025) integrate…

French frontier-model lab (Paris). Family includes Mistral Large 2 (123B dense, Jul-2024), Mixtral 8x22B (open-weights MoE), Mistral Small 3 (24B, open-weights Jan-2025), Codestral (code specialist…

OpenAI's foundation model family — GPT-5 (2025-08 flagship), GPT-4o (multimodal workhorse), o3/o4 reasoning series. Substrate for ChatGPT, the OpenAI API, Microsoft Copilot, Azure OpenAI, and many …

SF / Singapore foundation-model startup founded by ex-DeepMind / Meta / Google researchers (Dani Yogatama, Yi Tay, Qi Liu). Native multimodal models — Reka Core (~67B, frontier candidate), Flash (~…

Elon Musk's xAI frontier model — Grok 4 launched 2025-07. Trained on Colossus (100k+ GPU supercluster in Memphis). Deeply integrated with X/Twitter (real-time tweet data access). Grok 4 Heavy varia…

Open-source AI automation platform. Routes between short-term and long-term memory adaptively across any LLM provider; plugin system for storage backends. Memory managed at the instruction-manageme…

Agno (formerly Phidata). AgentStorage persists sessions to a DB; AgentMemory auto-classifies/store user preferences and conversation summaries. Single-line integrations with LanceDB, Pinecone, Weav…

ListMemory chronological context + teachable agents that vectorise corrections. Integrates with Mem0/Zep rather than building deep memory natively.

Per-plan vector-DB storage quota + LLMz autonomous engine (in-session working memory) + Knowledge Base (semantic search over uploaded docs). Long-term user memory persists across sessions.

Memory subsystem inside the CrewAI orchestration framework; integrates with Mem0 for the long-term tier.

dspy.History primitive — typed field holding messages: list[dict] that slots into any Signature . No persistent memory of its own; purely a structured context-injection contract. DSPy's optimisatio…

Visual LangChain canvas. LangChain-native memory nodes (Buffer, Buffer Window, Conversation Summary Buffer) backed by internal chat_message DB table. Buffer Window limits to last K interactions; Su…

Google's Agent Development Kit. Session services: InMemory / Database / VertexAI. Long-term: InMemoryMemoryService or VertexAIMemoryBankService (uses Vertex AI Memory Bank Preview for persistent cr…

Haystack 2.x: InMemoryChatMessageStore (experimental) + ChatMessageRetriever + ChatMessageWriter components in declarative pipelines. Legacy: ConversationSummaryMemory with LLM-compressed rolling s…

Memory as Network State — shared key-value store owned by the multi-agent network, readable by router / agents / tools. Persisted durably via Inngest infrastructure. Memory writes decoupled from re…

Agent component has built-in chat memory enabled by default, keyed by session ID. Custom session IDs isolate per-user history. Bundled Mem0 + Redis integrations for external memory layers.

Distinct from LangMem. Built-in checkpointer saves graph state per super-step (short-term, thread-scoped). Store System adds long-term hierarchical key-value memory across threads with optional vec…

First-class memory module inside LangChain — explicit cognitive-science-style typing of memories.

Saves high-signal facts (scheduling preferences, communication patterns, recurring rules) selectively rather than raw transcripts. Memory entries editable, injected into every prompt call, persist …

Two-tier model: ChatMessage FIFO queue with overflow flushed into long-term Memory Block objects. Memory class replacing older ChatMemoryBuffer.

TypeScript-first agent framework. Default memory: LibSQL store + on-device fastembed-js for semantic recall.

Low-code workflow platform. AI Agent node ships with pluggable memory sub-graph: Simple Memory for development; Postgres/Redis for production; vector store nodes (Qdrant, Pinecone, MongoDB Atlas) f…

OpenAI's agent SDK. Sessions for working context within an agent loop; RunContextWrapper for structured persistent state. April 2026 update added categorised long-term memory: project / user / poli…

PraisonAI multi-agent framework's built-in memory: per-agent short-term + cross-agent long-term, with Mem0 as the default long-term backend.

Pydantic-AI ships no native persistent memory; agents start fresh each run. Long-term memory typically added via third-party integration (e.g. Hindsight) using tools + auto-injected memory instruct…

Memory exposed via Anthropic SDK's BetaAbstractMemoryTool when Anthropic provider is active. No built-in persistent storage in the framework itself. Other providers require external memory (Hindsig…

Two memory scopes: conversation metadata (KV per-conversation) and persistent memory (cross-session). Enabling metadata auto-adds Add/Read conversation metadata tools.

Voice agents. Short-term context within call native; long-term memory via webhook integration to CRM / external DB. Cross-agent memory sharing in multi-agent setups.

Open-source visual AI programming environment. Memory is a first-class graph node: stores text block, supports stable IDs (saving same ID overwrites enabling updates), injects stored text into down…

Microsoft Semantic Kernel's memory plugin ( Microsoft.SemanticKernel.Plugins.Memory ). Plug-in approach to long-term memory in the SK agent framework.

Hugging Face barebones agent framework. Memory lives as a Python list; no built-in persistence. Step callbacks let you mutate memory between steps. Persistence is left to the user — community propo…

Wraps conversation memory behind an Advisor abstraction. MessageWindowChatMemory default 20-msg sliding window (preserves system messages); pluggable ChatMemoryRepository backends (in-memory, JDBC,…

Connects enterprise knowledge sources (SharePoint, Confluence, Notion, Drive, DBs) as memory layer with access controls, versioning, citations. Agents query at runtime; no separate session/cross-se…

AWS code-first agent framework. Multi-LLM (Bedrock, Anthropic, Gemini, OpenAI). Memory via Mem0 integration (ElastiCache for Valkey + Neptune Analytics) or Bedrock AgentCore Memory. Local default u…

Voice agents. Externalises memory from LLM into structured call state — variables rather than model recall — reducing hallucinations. Squad mode passes context across multi-agent assistant hand-off…

Chatbot platform. Tracks last 25 turns (configurable up to 100) as labelled assistant/user transcript stored in vf_memory string variable. Optional indefinite chat persistence across browser sessions.

Commercial platform built on Ray — distributed training, fine-tuning, serving. Anyscale Endpoints provides hosted OSS inference; ray cluster is the substrate underneath OpenAI's training stack.

Inference platform for ML models — Truss package format, Chains for multi-model workflows, autoscaling GPU/CPU serving. Targeted at production teams shipping LLM/diffusion endpoints.

Fast inference platform for OSS LLMs — custom Cuda kernels, speculative decoding, multi-LoRA serving. Targets latency-sensitive production use cases.

BerriAI's open-source LLM gateway — unified OpenAI-format API for 100+ providers; LiteLLM Proxy adds budgets, fallbacks, observability, rate-limiting.

Serverless cloud for AI/ML — Python-decorator workflow defines GPU/CPU functions deployed to managed infra; widely used for training jobs, batch inference, and agent backends.

Inference cloud spun out of OctoML — acquired by NVIDIA in 2024. Originally TVM-compiler heritage; offered API for OSS LLM/diffusion serving.

Unified API + marketplace for 300+ models — pay-as-you-go credits, automatic fallback, model arena. Used as default LLM router by many JS/TS agents.

AI gateway + observability — load balancing across 200+ LLMs, retries, caching, eval, prompts. Open-source Gateway + commercial Portkey Cloud.

Hosted-API platform for OSS models — Cog format packages models with their environment; thousands of community models behind a single REST API.

GPU-cloud platform — Pods (long-running), Serverless (per-request), and Hub (community templates). Popular for fine-tuning and inference of OSS models.

Fast LLM serving framework with RadixAttention (cache reuse across requests with shared prefixes) + structured programming primitives. UC Berkeley.

NVIDIA's TensorRT extension for LLM inference — compiles transformers to optimised kernels for Hopper/Blackwell. Used in NIM and downstream deployments.

Hugging Face's production inference server for LLMs — continuous batching, FlashAttention, quantization (bitsandbytes, GPTQ).

Cloud platform for OSS LLM inference and fine-tuning — serverless inference, dedicated endpoints, fine-tuning API, training cluster.

Open-source LLM inference engine with PagedAttention — high-throughput batching, paged KV-cache. UC Berkeley origin; de-facto OSS inference stack.

RDF triple/quad store with RDFS++ / OWL reasoning. v8.4 (May 2025) added an NLQ interface.

Vector index on graph nodes queryable via openCypher. Mem0 integration GA 2025; Cognee integration for agentic RAG. Combines semantic recall with multi-hop traversal in one managed service.

openCypher over Postgres. Pairs with pgvector for graph + vector hybrid retrieval. Azure Database for PostgreSQL ships AGE with ai_extension LLM functions for entity extraction directly in SQL.

HybridGraphRAG combines vector search, graph traversal, full-text in one AQL query. ArangoGraphML for ML pipelines; LangChain integration.

Vector indexing on any node. Google Gen AI Toolbox integration; LangChain agent orchestration.

Auto-built knowledge graph from web crawl. 1T+ facts, 10B+ entities. GraphRAG-fine-tuned model based on open-source Llama 3.3.

Native GraphRAG SDK. Graph-memory backend for Mem0; MCP server via Graphiti integration; context graphs for long-term agent memory.

Embedded graph DB with built-in vector index + full-text search. MCP server packages for Claude (sub-3ms recall). Offline-first, single-file deployment. Community fork adds concurrent multi-writer …

In-memory GraphRAG engine. Memgraph 3.0 adds native vector search alongside graph traversal. LangChain toolkit.

v5.2 adds graph-vector-text hybrid retrieval natively. MCP server for Claude/GPT agents. GraphRAG in three lines of code.

Property graph DB with Cypher; native vector search; Aura Agent + MCP server for graph-as-memory. $100M GenAI investment in 2025.

RDF store with real-time inference. Ontotext merged with PoolParty in 2025 to form Graphwise.

Proprietary operational ontology — semantic objects + kinetic actions wired to microservices. Not open-standards KG; closed platform.

Taxonomy and knowledge organisation. Taxonomy Advisor (LLM-powered). SKOS-based.

RDF-native semantic platform with SHACL validation, virtual graphs, and the Voicebox AI assistant. $32.5M funding. Pitches itself as a governed-knowledge layer.

Parallel graph DB with GSQL/GQL. CoPilot for entity extraction. Free community edition.

Ontology-native semantic governance suite. SHACL + AI Copilots for taxonomy management.

Google DeepMind's AndroidWorld — 116 hand-crafted tasks across 20 real Android apps (Calendar, Files, Markor, Tasks, etc.). Evaluates LLM agents controlling a real Android device via accessibility …

Sample-efficiency benchmark protocol on Arcade Learning Environment (ALE) — agents limited to 100k environment steps (~2 hours of game-play) before evaluation. Introduced by SimPLe (Kaiser et al. 2…

Extends 20 bAbI reasoning tasks (fact chaining, induction, deduction, counting, lists) by scattering relevant sentences across PG-19 book text. Scales to 10M tokens. Tests whether models can chain …

BALROG (Benchmarking Agentic LLM and VLM Reasoning on Games) — 2024 benchmark suite covering NetHack, BabyAI, Crafter, MiniHack, TextWorld, Baba Is AI. Specifically targets long-horizon, partial-ob…

Anthropic's public Twitch livestream of Claude (3.7 Sonnet → Claude 4 Opus) playing Pokemon Red. Agent uses screenshot vision + Game Boy controller actions + a scratchpad memory file the model can …

Danijar Hafner's 2D Minecraft-inspired benchmark — open-world survival with 22 hierarchical achievements (collect wood → make pickaxe → mine stone → ... → collect diamond). Procedurally-generated 6…

Meta AI's Habitat 3.0 — extends Habitat 2.0 with avatar humans, social tasks (navigate-while-following, rearrangement-while-not-blocking), and human-robot collaboration evaluation. Foundational for…

DeepMind's Hanabi Learning Environment — a benchmark for cooperative multi-agent agents under partial observability and theory-of-mind. Players see each other's cards but not their own; success req…

Princeton NLP. 7 task categories — RAG with real retrieval, cited generation, summarisation, re-ranking, classification, many-shot ICL, long-document QA — at lengths up to 128K. Model-based evaluat…

Full-day hybrid workshop held April 27, 2026, Rio de Janeiro. Scope: episodic and semantic memory, working memory, parametric knowledge, KGs, vector DBs, retrieval pipelines, context management, lo…

300 items across three paradigms: procedural memory (one-shot skill retention after interference), priming (theme-driven response biases), classical conditioning (CS-US exposure shaping first decis…

Tsinghua / OpenBMB. First benchmark with average input >100K tokens. 12 tasks across English + Chinese: long novel QA, code debugging, math in long context, summarisation, retrieval. Realistic docu…

Conversations spanning up to 35 sessions / 300 turns / ~9K tokens, built from LLM-agent personas grounded on temporal event graphs and verified by human annotators. Tests QA (single-hop, multi-hop,…

Tsinghua THUDM. 21 datasets across 6 categories (single-doc QA, multi-doc QA, summarisation, few-shot ICL, synthetic, code completion) in English + Chinese. ~6.7K words English / 13.4K chars Chines…

5 core memory abilities — info extraction, multi-session reasoning, temporal reasoning, knowledge updates, abstention — via 500 curated questions in synthetic chat histories. Two scales: LongMemEva…

Renmin University. Evaluates LLM-agent memory mechanisms across three axes — effectiveness (accuracy), efficiency (LLM call count), capacity (degradation as memory grows). Distinguishes factual vs …

QA tasks across 7 domains / 36 risk types plus workflow tasks adapted from 20 Agent-SafetyBench environments with noisy tool returns; both use mixed benign-and-misleading memory pools in multi-roun…

HUST. Reconstructs existing datasets into incremental multi-turn dialogue chunks fed in time order. Plus EventQA (accurate retrieval) and FactConsolidation (selective forgetting). Targets retrieval…

Stanford Digital Economy Lab et al. Benchmarks memory in realistic Memory-Agent-Environment loops where tasks span multiple sessions and subtasks are explicitly interdependent — agents must distill…

NVIDIA + Caltech open-ended embodied-agent benchmark on Minecraft. 3000+ programmatic + creative tasks across survival / harvest / tech-tree / combat / creative. Includes an internet-scale knowledg…

MineRL is a research project / competition series using Minecraft as a long-horizon RL environment. The flagship Diamond Challenge tasks an agent with obtaining a diamond from raw world — requires …

DeepMind / FAIR's NetHack Learning Environment — a Gym-compatible interface to NetHack, the canonical roguelike. Massive procedural state space, permadeath, deep item / skill interactions, ASCII pa…

Places a known sentence ("needle") at varying positions inside a long irrelevant document ("haystack") and prompts the model to retrieve it. Tests context lengths from 2K to model max. Greg Kamradt…

OpenAI's Gym Retro — wraps thousands of classic arcade games (Sega Genesis, NES, SNES, Atari 2600, Game Boy) as Gym RL environments via libretro emulators. Released 2018; powered the OpenAI Retro C…

Community benchmark suite for LLM agents playing Old School RuneScape, a sandbox MMORPG with deep partial-observability, skill-grinding loops, and economy interaction. Tests long-horizon planning (…

HKU NLP + Salesforce's OSWorld — full-desktop agent benchmark with 369 tasks spanning real applications (LibreOffice, GIMP, Chrome, VS Code, Thunderbird, Files) across Ubuntu, Windows, and macOS. T…

Pokemon Red has emerged as a community + academic benchmark for long-horizon agentic AI. Notable instances: Peter Whiddy's PokemonRedExperiments (RL via PPO + curiosity, viral YouTube series 2023);…

OpenAI's 16-game procedurally-generated benchmark for evaluating RL generalisation. Unlike Atari (fixed levels memorisable), every PROCGEN episode uses a different procedural seed → test-set genera…

NVIDIA. Extends NIAH with multi-needle retrieval, multi-hop entity tracing, aggregation, QA — configurable for arbitrary length / difficulty. Evaluated 17 long-context models on 13 tasks.

Microsoft Research's SmartPlay — LLM-agent benchmark on 6 games (Bandits, Rock-Paper-Scissors, Tower of Hanoi, MessengerEnv, Crafter, Minecraft). Tests 9 distinct capability axes including object u…

Microsoft Research's TextWorld — a sandbox for training and evaluating RL agents on text-based games. Procedurally generates interactive-fiction games of tunable difficulty (map size, quest length,…

CMU's WebArena — 812 hand-crafted web tasks across 6 self-hosted reproductions of real websites (Reddit-clone Postmill, GitLab, OneStopShop, CMS, Map, Wikipedia). Released ICLR 2024. Distinct from …

Runtime enforcement targeting memory poisoning, unauthorised execution, identity spoofing per the OWASP LLM threat list. Visibility/control over MCP servers and locally-installed AI tools — the inf…

Applies guardrails at three points: (1) before write to vector DB, (2) before query reaches embedding model, (3) before response. Detects malicious instructions in stored memory before retrieval; s…

Targets the supply-chain / lineage layer rather than runtime memory writes. Model Genealogy tracks training/fine-tuning/modification history — catches poisoning baked in during training. AIBOM gene…

Screens every prompt, response, and retrieved document for indirect prompt injection — primary vector for memory poisoning. Treats memory as untrusted: anything written to or read from memory is ad…

Commercial Mem0 ships SOC 2 / HIPAA, zero-trust access controls, BYOK encryption, real-time monitoring, audit logs, workspace governance as defaults. OpenMemory is the local self-hosted variant (Do…

Cross-Model Verification Kernel (CMVK) requires majority-voting agreement across multiple model calls before a memory-influenced action. Agent OS package intercepts every action (memory included) a…

Open-source runtime defense. Enforces YAML policies on every memory read/write. SHA-256 baselines detect tampering, injection, sensitive-data leakage, protected-key modification, rapid-change anoma…

Bi-temporal knowledge graph preserves full lineage of every stored fact — every memory write traceable and auditable by graph schema, not bolted-on logging. SOC 2 Type II; HIPAA BAA available; BYOK…

When Mem0 is connected, gains Memory Operation Timeline, Search Analytics, Memory Growth tracking, Error Tracking per memory call. Standalone, records context at each step but doesn't analyse memor…

Treats memory as first-class in multi-agent tracing. Luna-2 SLMs (3B / 8B) scan every interaction for intent drift and belief drift; 20+ checks at sub-200ms latency. Catches when agent A's view of …

Memory module reads/writes captured as named spans. Trace Log View concatenates every agent step including memory ops. Agent Graphs (GA 2025) infer graph structure from observation nesting; session…

Memory reads, vector DB retrievals, state changes are distinct span types in traces. RAG eval separates retrieval quality (context precision) from generation quality (faithfulness). Dataset version…

Only tool found that scans the persistent memory layer on disk rather than runtime tracing. Detects injected instructions, obfuscated payloads (zero-width Unicode, base64, homoglyphs, hex), C2-patt…

Founded 2022 by ex-OpenAI / Google researchers (David Luan, Kelsey Schroeder). Built ACT-1 / ACT-2 multimodal action transformers for computer-use. **Acquired by Amazon June-2024 (acqui-hire)** — c…

Burr is a state-machine framework for LLM agents from DAGWorks (commercial company behind Hamilton dataflow framework, $4M seed). State-machine abstraction instead of DAG / ReAct — more debuggable …

Commercial Enterprise tier of the CrewAI OSS multi-agent framework. Adds managed cloud runtime, RBAC, audit logging, SLA support, on-prem option. OSS CrewAI is in the Framework-embedded memory sect…

Founded 2017 as Generally Intelligent; rebranded Imbue 2023. Building foundation models for reasoning agents that write/edit code. Founders Kanjun Qiu + Josh Albrecht. $200M Series B 2023 ($1B valu…

Red Hat / IBM-stewarded open-source framework for community-driven LLM alignment + fine-tuning. Method paper 'LAB: Large-Scale Alignment for Chatbots' (Sudalairaj et al, IBM Research 2024). Built o…

Lindy is a no-code multi-agent automation platform from Florent Crivello (ex-Teleport, ex-Twitter). Targets consumer / SMB segment with email triage, meeting scheduling, CRM-update agents. $50M Ser…

Microsoft Research's low-code GUI for designing and deploying multi-agent workflows on top of the AutoGen framework. AutoGen v0.4 (Jan-2025) re-architecture moved to actor-based async runtime. Free…

Web-browsing agent API platform. Founded 2023 by Div Garg (Stanford); $30M+ raised. Provides 'Agent API' for autonomous web tasks (booking, shopping, research). Distinct from browser-agent memory s…

Phidata rebranded to Agno in 2024. OSS Python framework for building multi-agent systems with memory, knowledge, tools, reasoning. Lightweight alternative to LangChain. ~25k+ stars combined Phidata…

Founded 2024 by Misha Laskin and Ioannis Antonoglou (both ex-DeepMind, AlphaGo / Gemini RLHF). Building autonomous coding agents — Asimov is their code-research agent. Raised $130M Series A March-2…

Sema4.ai is the AI-agent rebrand of Robocorp (RPA / Python automation OSS). Funded by Benchmark, Canvas Ventures, Slow Ventures. Founded 2019 (as Robocorp); rebranded 2024. Positioned as enterprise…

YC W21 startup providing managed hosting + deployment for LLM agents — 'one-click deploy' for LangChain / agent code. Founded by Ted Benson (ex-Google Brain). Pre-seed + seed funded YC + Twin Ventu…

Tags + bidirectional links + task/calendar integration. "Task Score" priority matrix. AmpleAI plugin supports OpenAI, Anthropic, Gemini, Grok, DeepSeek, Ollama.

Folder + tag hierarchy; iCloud sync. Writing Tools (proofread / rewrite / summarise) across all note content; audio transcription + AI summary; smart search; reminders from to-dos.

Object-graph (people, books, projects as typed objects); daily notes; bidirectional links. AI chat grounded in current note; AI property auto-fill; in-editor writing assistant; pattern-finding acro…

Hierarchical block-based documents + nested pages (no native graph). In-document AI assistant: summarise, rewrite, translate, Q&A over docs, proofread.

Visual whiteboard canvas + nested cards + bidirectional links + journal. AI chat with cited sources inside whiteboards; supports OpenAI / Gemini / Anthropic backends.

AI memory and research assistant. Resurfaces pages, highlights, and notes while browsing; saves context, recalls sources, generates summaries.

Open-source local-first outliner. Notes stored as Markdown / Org-mode on disk. AI integration via community plugins (e.g. logseq-ai for chatting with the graph). Graph view + backlinks; daily-journ…

Note-taking with no folders; AI-chat over notes. Focused on connections between notes rather than hierarchy.

Windows-native screen-recording memory. Continuously indexes activity across apps, documents, browsing; timeline UI for revisiting work. Distinct from Copilot Memory.

No tags, no folders, no manual organisation. Automatic image tagging, OCR search inside screenshots/memes, smart auto-grouping (Smart Spaces), semantic search. Privacy-first.

Notion's first-party AI. Reads up to 50 pages per query (as of Jan 2026); Custom Instructions persist brand voice / team structure / preferences. Context follows you across model swaps. Notably not…

Obsidian community plugin. Generates embeddings locally with a bundled model (TaylorAI/bge-micro-v2, 384-dim); chat with notes + see semantically related links. No API key required for local; suppo…

Personal AI product. Builds a personal language model from your captured memories, messages, and content; queryable as "your AI."

Networked-thought note app with AI features. Apple-ecosystem-native; backlinks + daily notes + AI synthesis.

Hierarchical outline + bidirectional links + spaced-repetition flashcards. AI auto-generates flashcards/quizzes from notes; AI summaries; fill-gap questions.

Auto-tagged flat notes; semantic clusters (no folders). "Skai" personal AI surfaces related notes, semantic search, cross-tool integration (Drive, Slack, Calendar). ADHD-focused.

Infinite whiteboard + bidirectional-linked cards (spatial + networked hybrid). AI assistant for brainstorming, summarisation, rewriting selected cards.

Flat notecards + tags + collections; bidirectional references. AI tag recommendations, grammar suggestions, Vision (notecard editing assistant), Collection Creator.

Outliner + structured-data hybrid (Roam/Logseq + Notion-database influences). "Supertags" turn freeform nodes into queryable databases; Tana AI invokes on any node to summarise, generate action ite…

Tencent's knowledge-management productivity AI, powered by Hunyuan. Copilot mode adds a four-module persistent memory architecture: Soul (agent persona), User (profile), Memory (long-term declarati…

Memory tool API + consumer "Memory" feature + Auto Dream consolidation. Document-as-memory rather than extracted-fact-as-memory.

Mounts memory as files on a managed filesystem; Claude reads and writes via the same bash/code-execution tools used for agentic tasks. Unit of storage is a file — inspectable, editable, exportable …

AWS managed memory service for agents running on Bedrock AgentCore. Native integrations with LangChain/LangGraph and Strands. Drives long-running MCP servers + agent workloads.

Managed persistent memory service for AI agents on Cloudflare's edge. Beta-launched April 2026. Sits alongside Workers AI / Vectorize as the memory tier of the Cloudflare agent stack.

GA at Google Cloud Next 2026 (2026-04-22). Vertex AI rebranded as Gemini Enterprise Agent Platform; Memory Bank now "Agent Platform Memory Bank" with new continuous event-streaming + automated memo…

Personal Context setting + Personal Intelligence packing. Memory import from ChatGPT/Claude (Mar 2026).

Extracts and persists facts from past conversations; applies them to future responses. Per-memory deletion UI; user can also clear all.

Production memory infrastructure powering LinkedIn's Hiring Assistant and other GenAI apps. Three-layer (episodic/semantic/procedural) shared memory substrate across multi-agent systems; recent-con…

Stores facts from one-on-one chats and shares them across Instagram, WhatsApp, Messenger, Facebook when accounts are linked via Accounts Center. The widest cross-surface memory scope of any consume…

Microsoft 365 Copilot memory stored in user's Exchange mailbox with compliance alignment. GA Jul 2025.

Mistral's Agents API provides persistent conversation state across turns and sessions; runtime maintains branching conversation trees. Memory is implicit in the stateful session model rather than a…

Le Chat Memories — graph-based, context-aware. Stores user preferences and past decisions; memories saved explicitly (user instruction) or inferred automatically. Coupled to "Work Mode" agentic layer.

First-party memory in ChatGPT; auto-extracts persistent facts about the user across sessions. Apr 2025 expansion to chat-history referencing.

Auto-infers facts from conversations and surfaces them in future responses (Pro subscribers). Cross-model: memories persist when user switches underlying model mid-session. Enterprise Memory adds o…

Google Cloud managed memory service (Preview). Used by ADK's VertexAIMemoryBankService for cross-session persistent memory. Intelligently stores and retrieves memories from past user interactions.

Optimises context compression for long-horizon LLM agents.

Leverages cross-domain experience for agentic problem solving.

Open agentic framework that uses computers like a human.

Workflow-based memory framework component.

Agent that maintains a skill library and reconfigures its toolset per task — procedural memory of successful sub-task solutions.

Towards an efficient self-evolving agent system.

Long-horizon web agents with proactive context management.

Learning unified long-term and short-term memory management.

Extracts, stores, adapts, and reuses structured plan templates from planning stages of agent applications. NeurIPS 2025 poster.

Life-long personalisation of LLMs.

Global Workspace Layer combining low-rank explicit memory + bottleneck attention + Hopfield-style associative memory. Decouples memory from the Hopfield network via learnable explicit memory. Outpe…

Generalist agent enabling scalable agentic reasoning.

Self-evolving generative agent for agent generation.

Null-space-constrained knowledge editing for language models.

Replaces MLP layers with sparse cross-layer transcoders to produce "replacement models" where activations correspond to interpretable features; traces attribution graphs showing which features caus…

Episodic observations as timestamped vertices; LLM extracts relational triplets into a semantic knowledge graph; episodic edges connect semantic nodes — structural equivalent of hippocampal index b…

Associative Recurrent Memory Transformer. Stores tokens in Hopfield-style energy basins for constant-time pattern completion.

Long-term memory module that learns to memorise the context at test time. Omega rule (sliding-window update over all past tokens), polynomial feature mappings, Muon optimiser. Generalises to "DeepT…

Builds web agents with human-inspired web-browsing actions.

Thought-augmented reasoning with LLMs. NeurIPS 2024.

Constructivist view of agentic memory for reading comprehension.

Context-Dependent Memory framework. Three-stage encoding (expert / short-term / long-term) + graph-structured context-dependent indexing. NAACL 2025 Industry.

Comprehensive survey. ACM Computing Surveys 2025. Anchor reference for the continual-learning-of-LLMs subfield.

Scalable multi-agent framework for Windows UI task automation.

Cognitive-inspired memory-organised RAG for stateful narrative reasoning.

Maintains recent states in full resolution while compressing older memories with learned compression functions. DeepMind.

Tool creation for disentangling abstract and concrete reasoning. EMNLP Findings.

Enhances dialogue consistency via dynamic structured memory and reasoning.

Open-ended evolution of self-improving agents.

General reasoning agent with scalable toolsets.

DIffusion As a Model Of eNvironment Dreams. Trains an RL agent inside a diffusion-based world model — denoising diffusion generates high-fidelity environment rollouts rather than discrete latents. …

Encodes entire document corpora directly into Transformer parameters; the model itself becomes the index.

Hafner / Yan. Solves control tasks by training agents inside a fast, accurate world model.

Hafner et al. General world-model agent. Symlog observation, KL balance, RMSNorm, Block GRU.

System 1: fast retrieval path that pulls compressed, relevant memories and conditions generation. System 2: slow reflection pass that evaluates outcomes and writes curated updates back through a co…

Test-time learning with adaptive memory.

Enhances lifelong model editing with mixture-of-LoRA. AAAI 2025.

Uses Bayesian surprise detection to partition the cache into meaningful episodes for coherent retrieval.

Efficient Memory-Augmented Transformer. Encodes millions of QA pairs into key-value memory; MIPS search returns hits in sub-millisecond latency. EMNLP 2022.

Multi-agent RL with semantically coherent episodic buffer; selectively promotes desirable transitions to avoid local convergence.

DeepSeek + PKU. Introduces "conditional memory" alongside MoE's conditional compute. 27B-param N-gram lookup store offloads static knowledge from the transformer; O(1) lookup; prefetched from host …

Pink et al. — position paper. Uses CLS theory to identify five structural properties of episodic memory absent from current LLM agents: single-shot acquisition, temporal context, source binding, fl…

Kirkpatrick et al. Soft quadratic constraint pulls each weight back toward old values, scaled by importance for prior tasks. PNAS 2017.

LLM agents are experiential learners. AAAI 2024.

Cross-task generalisation by collecting and abstracting successful trajectories into insights for future tasks.

Synthesised LLM multi-agent system with conceptual verbal reinforcement.

Continuous agent evolution via forward learning from experience.

Replay scheduler inspired by the Ebbinghaus forgetting curve, plus intensity-aware regularisation to control replay strength. Mitigates catastrophic forgetting in LLM continual learning.

Tracing hierarchical memory for multi-agent systems.

Separates offline memory (lightweight Memorizer highlighting key facts) from online runtime (Researcher agent that queries the full raw page-store at inference). Inverts usual approach — keep full …

Park et al. — landmark agent-simulation paper. Reflection + memory stream + retrieval enable believable agent behavior.

Stanford's iconic memory-stream architecture for believable simulacra — observations + reflections + plans + retrieval by recency/importance/relevance. UIST 2023.

DeepMind. Foundation world model trained unsupervised on internet gaming video. Generates frame-by-frame controllable virtual worlds from text / image / sketch prompts. Spatiotemporal video tokeniz…

Heavy-hitter oracle for efficient generative inference. NeurIPS 2023.

Heavy-Hitter eviction policy for KV cache — retains tokens with cumulative high attention scores; up to 5× throughput.

Hierarchical hindsight reflection for multi-task LLM agents.

Symbolic indexing + hierarchical memory refinement + structured episodic retrieval, inspired by hippocampal mechanisms. Nature Scientific Reports 2025.

Co-activation strengthens associative edges (Hebbian Distillation); separate consolidation pass periodically extracts semantic summaries from the episodic graph — directly modelling slow-wave hippo…

Re-implements ACT-R's retrieval probability model — temporal decay, prior-access frequency, stochastic noise — as scoring on top of a vector store. Spreading activation over semantic similarity add…

Subquadratic alternative to attention — implicit long convolutions + data-controlled gating. Stanford. Ancestor of StripedHyena.

Meta. Self-supervised image learning by predicting target-block representations from a context block in latent space — no hand-crafted augmentations or pixel reconstruction. Foundation of LeCun's h…

Meta-tokens are learned vocabulary entries that act as placeholders for recurring token subsequences — the model is fine-tuned to understand compressed input where meta-tokens stand in for their co…

Inducing programmatic skills for agentic task enhancement.

Google. "Leave No Context Behind." Compressive memory inside the attention block + masked local + long-term linear attention.

Rethinks long-horizon agents via Markovian state reconstruction.

AI21's hybrid Mamba + Transformer MoE model — 256k context window; open-weights. First production model with SSM-Transformer hybrid architecture.

Open-world multi-task agents with memory-augmented multimodal LLMs. TPAMI 2024 . (Distinct from any other JARVIS.)

Real-time LLM personalisation based on human feedback.

IBM Research. Attaches a fast-learning distributed episodic memory (hippocampus-inspired) to a frozen LLM, enabling one-shot knowledge updates without retraining. Supports selective forgetting + in…

Few-shot mobile GUI agent with unified demonstration benchmark.

Modular procedural memory for multi-agent LLM systems.

Lightweight, efficient memory-augmented generation.

Neuro-symbolic architecture for manufacturing decision-making. Extracts ACT-R's procedural / declarative / goal buffers as latent neural representations, then injects them into trainable LLM adapte…

Large memory models — explicit external memory module within transformer.

Shift-short attention + LoRA fine-tuning that extends LLaMA context length cheaply — 100k tokens on a single GPU. ICLR 2024.

Trainable SideNet module selectively retrieves relevant cached key-value pairs from memory banks during decoding.

Generative agents for low-cost real-time social interactions.

Extends MemoryLLM with scalable long-term memory.

Multi-graph agentic memory architecture for AI agents.

Selective state-space model — input-dependent SSM with linear scaling; competitive with transformers at smaller scales. CMU.

Structured State-space Duality — Mamba-2 architecture that unifies SSMs and attention; 2-8× faster than Mamba-1. ICML 2024.

Memory-Augmented Transformer Using Heterogeneous Knowledge Sources. Unifies unstructured paragraphs + semi-structured QA pairs into fixed-length type-agnostic neural memories. Jun 2024.

Learning memory construction via reinforcement learning.

Learns to synergise memory and reasoning for efficient agents.

Reshapes long-context LLM with multi-conv RL-based memory agent.

Preserves memory states across dialogue turns for coherent long-term interactions. (Cited in survey but the paper's exact arxiv/venue identifier was not recoverable in a quick search — likely a les…

Fine-tunes LLM-agent behavior without fine-tuning the underlying model.

Meta-evolution of agent memory systems. Includes EvolveLab — a unified self-evolving memory codebase distilling 12 representative memory systems.

External key-value stores with similarity-based cache management; trained with Memory Replay Backpropagation.

Treats past optimisation gradients as memory registers for procedural computation. Distinct from the 2010.06891 "Memformer" above. (Survey citation — distinct paper identifier not recovered.)

Weaves generative latent memory for self-evolving agents.

Already in catalog as the foundational MemGPT paper. Note: Letta is the productionised successor (cross-listed).

Intent-driven memory selection for goal-oriented multi-session agents.

Autonomously augments semantic representations of historical interactions to improve retrieval quality without manual curation. Validated on conversational recommendation, QA, and event summarisati…

Retrieval Causal Attention actively prunes less important cached entries.

Distils expert adapters for on-device memory systems.

Stores raw conversational episodes without LLM-driven extraction, indexing at sentence level; retrieval expands nucleus matches into neighboring-context clusters to address the embedding-dissimilar…

Co-pilot alongside reasoning agent. Builds dependency graph over reasoning steps; prunes invalid steps, folds completed sub-trajectories, preserves compact high-salience backbone under fixed contex…

Tunes LLMs to use memos for consistent long-range conversation.

Reifies the reuse of user-AI conversational memories. CHI/UIST.

Resolves fateful forgetting through human-inspired memory architecture. Distinct from MatrixOrigin's commercial Memoria.

Architecture with external retrieval memory in the transformer.

Concise interface for real-time memory augmentation. CHI.

Autonomous context curation for long-horizon agentic tasks.

Pretrained, plug-and-play memory for LLMs.

Meta. Replaces dense feed-forward blocks with trainable key-value layers using product-key lookup.

Three-tier memory model: implicit (weights), working (context), and explicit (compressed KV stored externally). 2.4B model competitive with 7B+ using explicit memory.

Language modeling with explicit memory. Landmark.

Enhances LLMs with long-term memory. Early influential paper.

Tsinghua THUIR. First benchmark covering both declarative and procedural memory for LLM systems. User-feedback simulation framework tests adaptive improvement. Distinct from Renmin's MemBench.

Augments a transformer with a fixed-size 1B-parameter memory pool distributed across all layers; new knowledge is injected by updating the pool parameters during inference, not via context extensio…

Explores agent procedural memory.

Closed-loop retrieval controller replacing single-shot retrieve-then-answer. Iterative router chooses between retrieve, reflect, answer; global evidence-gap tracker makes retrieval reasoning explic…

IBM Research. Memory-augmented LLM architecture for multi-hop reasoning. Iteratively re-reads a temporal latent memory with bidirectional GRUs until readout stabilises. NeurIPS 2024.

Self-evolving agents via runtime reinforcement learning on episodic memory.

Trains LLMs to reason, search, and manage memory via RL.

Optimises short-term memory management for dynamic tool calling.

Builds an LLM-driven hierarchical tree of facts; queries traverse the tree from root to leaves, balancing recall coverage with depth.

Multimodal memory system for lifelong learning agents.

Embedding-space memory consolidation — merges semantically similar memories without LLM arbitration to reduce storage cost.

Meta-programming for a multi-agent collaborative framework.

Manager stores compressed non-parametric search trajectories; Planner is a parametric memory agent producing search plans; Executor acts on plans. Planner evolves via test-time learning (on-the-fly…

Multi-agent memory system for LLM-based agents.

Language modeling with retriever-pretrained external memory.

Mixed memory-augmented generation for LLM applications.

Before test time, LLM reasons over unlabeled data and stores high-confidence intermediate thoughts as external memory. At test time, retrieves relevant stored thoughts to augment reasoning. No anno…

Boosts LLM agents with meta plan optimisation. EMNLP 2025.

Sakana AI. "An Evolved Universal Transformer Memory." Neural Attention Memory Models trained via evolution (overcomes non-differentiability of token-keep/discard).

Two complementary NAVER Labs Europe papers (Jan 2025). Provence fine-tunes DeBERTa as a sequence labeler to dynamically detect needed pruning amount per query and mask irrelevant context spans befo…

Self-organising agent memory inspired by cognitive science.

Pritzel et al. Agent stores past state-action-value tuples in a differentiable dictionary; recall via approximate nearest neighbour. ICML 2017.

First end-to-end pipeline translating natural language into executable Soar rules. Imports the full Soar production-rule cycle: working memory, operator proposal/application, chunking. LLM generato…

Interleaves sparse MoE layers with Mamba-2 layers; Mamba layers maintain a fixed-size recurrent state (constant memory during generation) rather than a growing KV cache. Attention layers use GQA wi…

Open agent social interaction simulations with one million agents.

Episodic simulation and memory for vision-language navigation.

Trains long-context models by simulating long positions with skip-wise positional encodings — avoids high training cost of true long sequences.

Planning- and retrieval-integrated memory for enhanced reasoning.

Synthetic strategy memory for proactive dialogue agents.

Pre-trains a large multi-task world model on offline data across many continuous-control tasks, then extracts task-specific policies by differentiating through the model with first-order gradient o…

Memory-enhanced model-based RL agent built on DreamerV3. Modifies S4 to handle long temporal dependencies. Keeps DreamerV3's generality.

Bridging memory retention and retrieval via reversible compression.

Efficient KV-cache compression through retrieval heads.

Google Research / DeepMind. Prompts LLM to segment long documents into "pages," compresses each to a short gist stored as episodic memory, re-reads original pages on demand at answer time. Pure pro…

Scales agent self-evolution with reasoning memory.

Bulatov / Kuratov / Burtsev. Adds memory tokens passed between segments — no architectural change to the Transformer.

Interactive generation of arbitrarily long text.

Language agents with verbal reinforcement learning.

Autonomous LLM-based agent for program repair.

Unlocks long-horizon search intelligence via context summarisation.

General read-write memory for large language models.

Retrospective LLM agents with policy-gradient optimisation.

Renormalization-group-based memory evolution for user profiles.

Uses tensor decomposition to exploit low intrinsic dimensionality and pixel correlation of stored exemplar images, achieving high compression while preserving discriminative information for class-i…

Local-first MCP architecture for RL coding agents that treats memory selection as a logged contextual decision process. Three MCP tools: issue_match (ranks candidates + records telemetry), issue_fe…

Combines forward-looking Prospective Reflection (multi-granularity summarisation into personalised memory bank) with backward-looking Retrospective Reflection (RL-driven retrieval refinement guided…

RWKV recurrent-LM family — linear-time, constant-memory inference; v7 ('Goose') is the SotA recurrent LM. Community-driven OSS, Linux Foundation hosted.

Social-network simulation system with LLM-empowered agents.

Self-evolving agents with reflective and memory-augmented abilities. Neurocomputing 2025 .

Continuous Thought Machines. Each neuron maintains its own history of incoming activations via private per-neuron weight parameters; the model operates along a self-generated "thought step" timelin…

Enhances LLMs with self-controlled memory framework.

Empowers LLMs with cognitive agency via active context management.

Novel activation approach for memory.

Self-evolving computer-use agent with autonomous learning.

Memory construction and retrieval for personalised conversational agents.

Encodes new experiences directly into the LLM's existing parameters by minimizing KL divergence between a teacher model (with context access) and a student model (without). Generates diverse QA pai…

Iterative self-feedback loop — LLM critiques its own output and revises; in-context experiential memory of the current task.

Dynamic regulariser driven by Wasserstein distance between task distributions; automatically relaxes or strengthens constraints in proportion to task similarity.

Sentence-graph memory for long-term conversational agents.

During prefill, value caches are offloaded to CPU and only low-rank pre-RoPE keys plus compressed landmarks and outliers are kept on GPU. During decoding, landmarks select chunk indices for key rec…

Web agents self-improve by discovering and honing skills.

LLM knows what you're seeking before generation. NeurIPS 2024.

Soft chain-of-thought for efficient reasoning with LLMs. ACL 2025.

Attention-sink discovery — keep the first few tokens always cached; arbitrary-length inference with no fine-tuning. ICLR 2024.

Memory is a dynamic graph; retrieval fires spreading activation across edges rather than vector similarity. Lateral inhibition suppresses competing activations; temporal decay modulates edge weight…

Stanford's Test-Time Training layers — RNN-like hidden state is itself an inner ML model that updates at test time on the current input.

Two-tier: fast compressed summary index (Tier 1) + immutable raw-log store (Tier 2), with explicit provenance pointers linking summary units to source pages. Runtime sufficiency router escalates to…

Language models teach themselves to use tools.

Unified tool retrieval and calling via generation.

Facilitates LLMs mastering 16,000+ real-world APIs.

Enhances multimodal agents with learnable tool capability memory.

Extends context through segment-level recurrence + caching of hidden states from prior segments. Foundational long-context architecture.

Encodes procedural expertise using SVD decomposition with dynamic expert-vector blending.

Google. Feedback attention loop enables the model to attend to its own latent representations — emerges as working memory. No extra weights; integrates with pre-trained models.

Desktop AgentOS with memory-driven task automation.

Meta AI. Predicts masked spatio-temporal video patches in latent representation space rather than pixel space. Implicit world model without generative decoding. Released for self-supervised video r…

Meta. Self-supervised video model. Predicts in latent space (no pixel decoder); supports understanding, prediction, and planning downstream. June 2025.

NVIDIA lifelong-learning agent for Minecraft — automatic curriculum + skill library + iterative prompting. Foundational procedural-memory paper.

Structures web-scale evidence with dynamic outlines for research agents.

Rethinks knowledge memory for lifelong model editing. NeurIPS 2024.

Kruijne et al. RL-learned gating policies open/close multiple working-memory slots independently. Handles hierarchical 12-AX task with multiple concurrent context levels; all-or-nothing slot updates.

RoPE-extension method for context-length scaling — combines NTK-aware interpolation with attention scaling, requires only short fine-tuning. Widely adopted in OSS long-context models.

Treats memories as atomic linkable notes — explicit nod to Zettelkasten knowledge management. Dynamic linking; retroactive memory revision.

Hierarchical "OS" with Storage / Updating / Retrieval / Generation modules. Short-term → mid-term via FIFO dialogue-chain; mid-term → long-term via segmented paging.

Self-organizing memory OS for structured long-horizon reasoning. Three-phase model: episodic, semantic, reconstructive.

Dynamic Memory Network + Hierarchical Memory Manager + Self-Improvement Engine. L0 raw embeddings, L1 contextual summaries, L2 high-level principles. NeurIPS 2025 (Scaling Environments for Agents w…

Lightweight hierarchical graph (CogniGraph) with entities and relations as semantic indexing layers. Incremental graph construction, fast updates, low-latency inference. Nov 2025.

RAG framework on top of a long-context memory model. Builds global memory once, generates contextual clues at query time. TheWebConf 2025.

Treats memory as an OS-managed resource with explicit allocation and process-like scoping. MemCubes unify parametric, activation, and plaintext memory.

Spatial-metaphor memory (wings / rooms / halls) on top of verbatim chunked storage. Independent analysis suggests the score is driven by verbatim storage + ChromaDB defaults rather than the palace …

Temporal-Hierarchical Memory Consolidation. 5-layer Temporal Memory Tree (segments → profiles); semantic-guided consolidation without fine-tuning; complexity-aware recall planning + gating. Jan 2026.

Neural long-term memory module that learns to memorise at test time. Uses gradient-of-loss as "surprise" signal; adaptive weight-decay forgetting. Three variants: MAC (memory-as-context), MAG (memo…

Smaller classifier LM predicts query complexity, then routes to no-retrieval / single-step / iterative retrieval as appropriate. NAACL 2024.

Meta AI. Jointly fine-tunes retriever + LM for few-shot tasks.

Applies AutoML to RAG: greedy algorithm evaluates and selects best-performing module at each pipeline stage (query expansion, retrieval, reranking, generation). Validated on ARAGOG (423 papers, 107…

Argues standard RAG breaks on bounded agent conversation streams (highly correlated, redundant spans). Decomposes the stream into semantic units, organises them into themes, then inverts that struc…

Single embedding model supporting dense, sparse, and multi-vector (ColBERT-style) retrieval simultaneously. 100+ languages, inputs up to 8192 tokens. Self-knowledge distillation across the three mo…

"Don't Do RAG." Preloads all relevant documents into context, caches the resulting KV state; at inference the model answers from cache with no retrieval step — eliminates retrieval latency + retrie…

Semantic chunking divides documents into coherent sections; LLM-based relevance scorer filters each chunk against the query before generation, pruning irrelevant chunks and reducing hallucination.

Lightweight retrieval evaluator scores docs and triggers different actions: use, supplement with web search, or discard. Decompose-then-recompose algorithm filters irrelevant content from retrieved…

Forward-Looking Active Retrieval. Uses model's low-confidence token predictions as signal to anticipate future information needs and retrieve proactively. Retrieves only when token probability fall…

Replaces chunk-retrieval with neuro-inspired generative memory. Operator maps observations to intermediate semantic structures; Reconciler integrates them into a persistent workspace with temporal/…

Retrieval-augmented fine-tuning for knowledge graphs on graph databases.

Modular graph-based RAG. Leiden community detection over knowledge graph; community summaries at multiple levels. DRIFT Search hybrid (Oct 2024).

Hippocampus-inspired RAG. LLMs + knowledge graphs + Personalized PageRank to mimic neocortex/hippocampus roles. NeurIPS 2024.

OSU follow-on to HippoRAG — adds online updates, recall-precision tradeoff calibration, and multi-hop reasoning improvements.

Multi-layer knowledge graph: higher layers hold coarse summary entities improving connectivity between similar lower-layer entities (HiIndex). HiRetrieval traverses across granularity levels.

LLM generates a hypothetical answer document; embeds that instead of the query. Closes the query-document semantic gap for retrieval. arxiv 2212.10496.

Uses model's generated output as a rich context signal for the next retrieval round rather than just the original query. Processes all retrieved knowledge as a whole, preserving generation flexibil…

Knowledge-Augmented Generation. LLM-friendly KG representation + mutual indexing between KGs and raw chunks + logical-form-guided hybrid reasoning. Handles numerical / temporal / multi-hop chains t…

Performs minimal up-front indexing; defers LLM use until query time. Being integrated into the main GraphRAG library.

Hong Kong U graph-RAG variant — dual-level retrieval (local + global) over an entity-relation graph built incrementally from corpus updates.

Entity-relationship graph + broader-theme retrieval. Incremental update support. Integrates with RAG-Anything for multimodal docs.

Groups related Wikipedia documents into 4K-token units. Reduces NQ corpus 22M→600K units, HotpotQA 5M→500K. Hands longer passages to a long-context reader (GPT-4o).

RAG variant with a long-context global memory model that drafts answer clues used to guide passage retrieval — improves recall in noisy multi-hop QA.

Uses transformer multi-head attention activations (not decoder-layer embeddings) as retrieval keys, yielding multi-aspect embeddings that independently capture different semantic facets for heterog…

Retrieves over an indexed graph by pruning candidate relational paths — reduces noise in graph-RAG vs flat path expansion.

Retrieves key relational paths from knowledge graph; flow-based pruning + reliability scoring; path-based prompting. Outperforms baselines on six datasets. arxiv Feb 2025.

UC Berkeley. Trains the model to ignore distractor documents and quote the relevant text in chain-of-thought.

Stanford recursive-summarisation tree over the corpus — clusters and summarises documents into a hierarchical index; retrieves at multiple levels. ICLR 2024.

Recursively embeds, clusters, and summarises chunks into a tree of abstractions. Retrieval traverses the tree at query time. ICLR 2024.

Foundational paper. Augments LM pre-training with a latent knowledge retriever — model retrieves and attends over Wikipedia-scale corpus during pre-training, fine-tuning, and inference. Backprop th…

DeepMind. Retrieval-Enhanced Transformer with chunked cross-attention over a 2T-token retrieval database. Frozen BERT retriever + differentiable encoder.

End-to-end RL-trained RAG. Token-level policy chooses when to retrieve and which mode (text / KG / hybrid via RRF). Two-stage Group-Relative Policy Optimisation (GRPO). No public repo at time of wr…

Trains an LLM end-to-end to emit reflection tokens that decide whether to retrieve, assess passage relevance, and critique output quality — all within a single model.

Smaller distilled specialist generates multiple RAG drafts in parallel from distinct document subsets. Larger generalist verifies in a single pass. Subsets formed by content-similarity clustering t…

Identifies optimal structure (table, catalog, algorithm, graph) per task; converts source documents into that structure at inference time and reasons over the structured form. Motivated by cognitiv…

Frames RAG as Turing-complete with explicit monitored state variables, enabling adaptive retrieval halting and convergence control. Medical-LLM case study.

Domain KGs + multi-stage web-page retrieval (sparse + dense) + LLM self-assessment of trustworthiness before finalising answer.

Norway / US humanoid robot company — Eve (wheeled bimanual) + Neo (legged humanoid, consumer-targeted, late 2025 / 2026 launch). Raised $100M Series B (Jan-2024, EQT + OpenAI Startup Fund); $40M Se…

Oregon-based humanoid maker; Digit is in production at GXO Logistics warehouses since Sep-2024 — claimed first humanoid in commercial service in the US. Spinout of Oregon State University. $400M+ r…

Apollo humanoid maker, spun out of University of Texas Human Centered Robotics Lab. Raised $350M Series A (Feb-2025; B Capital + Capital Factory); Mercedes-Benz manufacturing partner (Mar 2024). Pa…

Founded 2022 by Brad Porter (ex-Amazon Robotics VP + Scale AI CTO). Stealthier than the humanoid pack; raised $100M Series B Apr-2024 (General Catalyst). Targets warehouse + healthcare collaborativ…

Pieter Abbeel / Peter Chen-founded robotics FM company — built RFM-1 (Robotics Foundation Model, Mar 2024) for warehouse pick-and-place. Acqui-hired by Amazon Aug-2024 (founders + 25% of team to Am…

Figure AI — humanoid robot maker founded 2022 by Brett Adcock (ex-Archer Aviation). Figure 02 (Aug 2024) + Helix VLA model (Feb 2025) for full-body / hand control. Raised $675M Series B Feb-2024 (M…

Google DeepMind's Gemini Robotics + Gemini Robotics-ER (Embodied Reasoning) released Mar-2025. VLA models built on Gemini 2.0 backbone. Reference platform: Apptronik Apollo. The DeepMind line of ro…

Hugging Face's open-source robotics platform — datasets, baselines, training code for imitation learning + diffusion policy + ACT. Launched May 2024; growing rapidly with community contributions (H…

Open-source humanoid robot maker — K-Scale OS (Kos), Stompy and KBot humanoids. YC W24. Pitched as the 'open-source humanoid' alternative to Figure / Tesla — full hardware + software OSS.

NVIDIA Project GR00T (Mar-2024) — general-purpose humanoid foundation model + Isaac simulation stack + Jetson Thor compute. GR00T N1 model released open March 2025. NVIDIA's full-stack play: model …

OpenMind — open robot OS (OM1) + FABRIC interconnect, founded by Jan Liphardt (Stanford). Pitched as the 'Android of robots': any humanoid + any LLM + the same OS layer. Raised $20M Series A (Pante…

Founded 2024 by Sergey Levine (UC Berkeley) and Karol Hausman + others (ex-Google Brain Robotics). Building general-purpose robot foundation models (π0 released Oct-2024; π0.5 in 2025) trained on c…

MIT spinout focused on truck-unloading robots — autonomously empties truck trailers. Raised $50M Series B (Mar 2024, Tiger Global + others). Narrow focus but production deployments at major logisti…

Vancouver-based humanoid maker; Phoenix robot powered by Carbon cognitive architecture. Founded 2018 by Geordie Rose (ex-D-Wave + Kindred). Less well-funded than Figure / 1X / Apptronik but with a …

Toyota Research Institute's robotics group — published Large Behavior Models (LBMs) Sep-2023 + diffusion-policy / behavior-cloning at scale. Industrial-research lab with Toyota's manufacturing-data…

Search-as-a-service API for instant search experiences — NeuralSearch (vector) + lexical. Adopted by 12000+ companies including LVMH, Stripe.

Elastic's managed Elasticsearch cloud — vector search (ELSER reranker, dense vectors), full-text, observability, security. Public (NYSE: ESTC).

French open-source search engine — fast, typo-tolerant, vector search beta. Meilisearch Cloud commercial.

AWS-led fork of Elasticsearch 7 — Apache-licensed; k-NN plugin, OpenSearch ML, neural-search. Substrate beneath AWS OpenSearch Service.

Open-source Postgres extension for vector similarity — HNSW and IVFFlat indexes; widely deployed (Supabase, Neon, AWS RDS, Crunchy, Azure).

Hosted managed Qdrant cloud — multi-cloud single-tenant clusters; hybrid cloud private deployment; widely used as a memory backend (Mem0 default).

Open-source distributed search engine for logs/traces — cloud-native (object storage as primary tier). Acquired by Datadog 2024.

Site-search platform (formerly Sajari) — vector and lexical search for e-commerce / content sites. Smaller player vs Algolia/Coveo.

Supabase's managed pgvector — combined Postgres + vector store; popular default for OSS LLM apps. Supabase $80M Series C 2024.

Rust port of Lucene — embedded full-text search engine. Underlies Quickwit, Lnx, Bayard. Used in Meilisearch and many in-process search engines.

Open-source search engine focused on typo-tolerance and ease of use — Typesense Cloud commercial; rising Algolia alternative.

Yahoo Japan's open-source distributed vector search engine — built on NGT; Kubernetes-native; horizontal scaling.

Hosted Vespa.ai (originally Yahoo internal). Hybrid search + ML inference at billions-of-documents scale. Used by Spotify, OkCupid, Wix.

Managed cloud SaaS for Weaviate — serverless or dedicated; multi-cloud (AWS/GCP/Azure); enterprise hybrid deployment.

Commercial managed cloud for Milvus — multi-cloud (AWS/GCP/Azure); enterprise tier with private deployment.

"+1 for 'context engineering' over 'prompt engineering' … the delicate art and science of filling the context window with just the right information for the next step." The LLM is "a coworker with …

Endorses "context engineering" as a distinct discipline from prompt engineering — what agents do with their context window (routing, compression, tool output formatting, memory retrieval injection)…

"Application context tends to expand to fill the context limits supported by the model." Treats agent memory as a first-class unresolved design problem rather than a solved component; explicitly de…

Traces the shift from weights-as-capability to harness-as-capability; analyzes memory, skills, and protocols as three coupled forms of externalization and examines how they interact. Memory is defi…

Coding agents (Claude Code, Cursor) converge on the filesystem as their primary memory abstraction: conversation histories as searchable files, skills as files, retrieval via file search rather tha…

"Agent = LLM + memory + planning + tools." Maps cognitive memory types to LLM machinery: sensory (learned embeddings), short-term / in-context (the context window), long-term (external vector store…

Eight-quadrant classification grid across personal/system, parametric/non-parametric, and short-term/long-term axes. Bridges cognitive-science memory taxonomy to LLM architecture choices, less comm…

Six-module cognitive architecture (perception, world model, cost, memory, action, configurator). Memory is a distinct addressable module with its own update rules, separate from the world model tha…

LLM is "the kernel process of a new Operating System" orchestrating I/O, code interpreter, browser, and an embedding store for files / internal memory. Context window = RAM; embedding store = files…

Operationalises the kernel analogy: 256-core processor at 20 Hz (tokens/sec), RAM = 128K tokens of context, filesystem = Ada-002 embedding store. Memory capacity becomes a measurable architectural …

Three-layer directory: raw/ (immutable sources), wiki/ (LLM-maintained compiled pages), CLAUDE.md (schema). LLM reads sources and integrates them into the wiki — updating pages, resolving contradic…

Argues that companion AI knowledge systems must treat memory as a continuous metabolic process rather than a static store. Proposes a three-tier store (raw buffer, active wiki, cold memory) governe…

Agent persistent memory is an attack surface: an adversarial document in the environment can instruct the agent's memory tool to write malicious facts, poisoning the long-term store. The more capab…

On-policy RL is naturally more resistant to catastrophic forgetting than supervised fine-tuning — its online nature biases learning toward low-distribution-shift updates. Proposal: RL post-training…

Hugging Face's abstraction over distributed-training backends (DDP, FSDP, DeepSpeed, Megatron). Minimal code change to scale a PyTorch script.

Open-source data-quality platform for LLM training data — human-in-the-loop labelling, RLHF data collection, dataset curation. Acquired by Hugging Face 2024.

AWS's managed ML platform — Studio, Training, HyperPod (large-scale clusters), Endpoints. Industry-standard for enterprise AWS-hosted training.

Open-source LLM fine-tuning library — single YAML config drives full / LoRA / QLoRA / DPO training across most popular base models.

Microsoft Azure's managed ML platform — pipelines, AutoML, Designer, prompt flow (LLM). Heavy integration with Azure OpenAI and Azure AI Studio.

Open model that generates instruction-tuning data conditioned on unstructured text — adapts pretrained LLMs to specialised tasks without manual labels (Brown Univ).

Distributed PPO/IMPALA reference implementations from CleanRL author — research-friendly, reproducible.

Open-source ML experiment tracking, orchestration, and data management. Allegro AI commercial; widely used in research labs.

HPC-AI Tech's distributed training framework — ZeRO + tensor/pipeline parallel + heterogeneous training (CPU offload). HPC-AI raised $50M.

ML experiment tracking + LLM evaluation platform (Opik for LLM eval). Comet competes head-to-head with Weights & Biases.

Open-source training library from MosaicML (Databricks) — Trainer + recipes for fast convergence; underlies the MPT model training stack.

Data orchestrator with first-class ML pipeline support — software-defined assets model; competes with Airflow/Prefect for ML pipelines.

Microsoft's distributed-training library — ZeRO optimizer states/grads/params partitioning; ZeRO-Inference for serving. Foundation under many OSS LLM training stacks.

Argilla's framework for synthetic data generation — pipelines of LLM steps that produce SFT/DPO training data; integrates with Hub.

Stanford method that converts RLHF into a supervised-learning loss over preference pairs — no separate reward model needed; widely adopted alignment recipe.

Open-source data and ML model version control — git-like commands for datasets/models stored in cloud blob storage. Foundational tool in MLOps stack.

Kubernetes-native workflow engine — type-safe Python tasks; Lyft origin; Union AI commercial. Strong adoption in regulated industries.

Native PyTorch fully-sharded data parallel implementation — Meta's response to DeepSpeed ZeRO; FSDP2 in 2024 with improved API.

Google Cloud's managed training service — TPU and GPU clusters, hyperparameter tuning, model registry. Part of Vertex AI umbrella.

DeepSeek-Math's GRPO — group-relative advantage estimation replacing PPO's critic. Used in DeepSeek-R1 reasoning post-training; widely adopted in OSS reasoning RLHF.

Library for accessing and processing 100k+ public datasets on the Hub. Apache Arrow-backed; default dataset layer for the open ML ecosystem.

Open-source data-labelling platform supporting text, image, audio, video, time-series. Commercial Enterprise edition + Cloud (HumanSignal).

Enterprise data-labelling platform — annotation tools + 'Frontier' offering (expert RLHF labelers + Boost compute). Pivoting toward frontier-lab data services.

Git-like branching, commits, and rollback for object-store data lakes (S3/GCS/Azure). Treeverse commercial; OSS + Cloud edition.

Unified efficient fine-tuning of 100+ LLMs — SFT, DPO, PPO, KTO, ORPO with LoRA/QLoRA/full; CLI + WebUI. Popular in the Chinese OSS scene.

Microsoft Research's Low-Rank Adapters method — trains rank-r matrices added to attention weights; 10000x parameter reduction. Foundation of all adapter-style fine-tuning.

PyTorch-library form of Megatron-LM — distributed training building blocks (tensor/pipeline/expert parallel) packaged for embedding into custom training stacks.

NVIDIA's reference framework for training large transformer models with tensor, pipeline, sequence, and expert parallelism. Underpins many frontier-lab training stacks.

Databricks-incubated open platform for ML lifecycle — tracking, projects, models, model registry. Standard ML-ops baseline. MLflow 3.0 (2024) added LLM-eval features.

NVIDIA's open synthetic-data generation pipelines (HelpSteer2, NemoSkills, Nemotron-4 340B). Used by frontier labs to bootstrap RLHF reward modeling data.

Distributed RLHF training framework built on Ray + DeepSpeed + vLLM — supports PPO/DPO/REINFORCE++/GRPO. Targeted at 70B+ models with efficient sample throughput.

Commercial platform behind Metaflow (originally Netflix OSS) — ML workflow orchestration with managed compute, focused on data scientists shipping production pipelines.

Data-versioning and pipeline platform — Pachyderm File System tracks lineage across pipeline runs. Acquired by HPE in 2023.

Hugging Face's PEFT library — LoRA, AdaLoRA, IA3, Prefix Tuning, P-Tuning. Standard interface for adapter-style fine-tuning.

Python-native workflow orchestrator — widely used for ML pipelines and ETL. Prefect Cloud + open-source Prefect Core.

4-bit quantized LoRA fine-tuning — train 65B models on a single 48GB GPU. UW (Tim Dettmers).

Ray's distributed training library — multi-node PyTorch/TF jobs; integrates Ray Tune and Ray Serve. Used at scale by OpenAI/Uber/Anyscale.

Allen AI's research library for RL training of language models — NLPO algorithm, NLP-task RL benchmarks; predates the TRL/OpenRLHF surge.

Research codebase for RLHF reward modelling — pairwise / Bradley-Terry / pointwise reward heads; release of RewardBench-winning models.

End-to-end computer-vision platform — dataset hosting, annotation, training, deployment. Universe public-dataset library; widely used for production CV.

ETH RSL's lightweight PyTorch RL library — optimised for GPU rollouts in robotics simulation (Isaac/MuJoCo). Used in many legged-robot RL papers.

Data-labeling powerhouse for AI training — RLHF data, fine-tuning corpora, frontier-lab partnerships. Meta invested $14B in Scale in 2025 for a 49% stake; Wang joined Meta as Chief AI Officer.

Programmatic data labelling and weak supervision platform — built on Snorkel research from Stanford. Enterprise platform for high-volume data programming.

PyTorch implementations of RL algorithms (PPO, SAC, DQN, etc.) — research-grade reliability; widely used as a reference benchmark.

Premium human-labelling company specialised in RLHF data and difficult expert tasks — Surge prides itself on bootstrapped, no-VC; major frontier-lab supplier.

Pure-PyTorch library for fine-tuning LLMs — Llama/Mistral/Phi/Gemma recipes; native FSDP2 + memory-efficient training. From Meta PyTorch team.

Hugging Face's open-source library for fine-tuning and aligning transformer models with RLHF, DPO, ORPO, KTO and PPO trainers. The default RLHF stack on top of HF Transformers.

Fast and memory-efficient LLM fine-tuning library — claims 2-5× speedup, 60-80% memory savings on a single GPU vs HF baseline through custom Triton kernels.

ByteDance Volcano-Engine's open RLHF framework — flexible HybridFlow architecture; supports PPO/GRPO at large scale.

Experiment-tracking and ML-ops platform — log metrics, artifacts, datasets, sweeps. Acquired by CoreWeave in 2025 for $1.7B.

Open-source ML pipeline orchestration framework — connects sklearn/PyTorch/HF pipelines to MLflow/W&B/cloud infra. Pivoting toward LLM-ops.

Founded 2022 by Hasan Sukkar; AI sales agents Alice (outbound SDR), Mike (voice SDR), Jordan (RevOps). $50M Series B Oct-2024 (Benchmark + Andreessen Horowitz; $350M val). Among the most-discussed …

Founded 2013; ABM/intent platform with Revenue AI agents. $200M Series E Jan-2022 ($5.2B val; Blue Owl). 6sense Revenue AI agents launched 2024 — autonomous account research + email drafting.

Founded 2010; AI-driven market intelligence — semantic search over earnings calls, broker research, regulatory filings. $4B val Jun-2024 ($650M growth round; BDT + MSP). **Acquired Tegus Jun-2024 f…

Founded 2019 by ex-Yahoo CISO Karthik Kannan; AI-driven detection-engineering platform. $45M Series C Jul-2024 (Evolution + G Squared; ~$200M val). Forge is the AI agent that auto-generates SIEM de…

Founded 2023 by Tomer Avni + Matan Getz (ex-Microsoft); generative-AI security platform spanning red-teaming, DLP, and runtime. $7M seed Apr-2024 (Sequoia + Index Ventures).

Founded 2015 (rebranded from ZenProspect); GTM platform — prospecting database + outreach + AI. $100M Series D Aug-2023 ($1.6B val; Bain Capital Ventures + Sequoia). $200M+ ARR; Apollo AI launched …

Founded 2014; enterprise audit + GRC platform — 60%+ of Fortune 500 customers. **Acquired by Hg for $3B Jun-2024** (private-equity take-private). AI features added 2024 — auto-population of audit w…

Dublin / NYC founded 2018; GenAI security and content scanning. $23M Series A-1 Aug-2024 (Paladin Capital + Lockheed Martin Ventures). Pivoted from defence ML to enterprise GenAI guardrails 2023-24.

Cross-listing: Causaly already exists in Scientific / research AI memory. This row adds the use-case agent harness framing: London-based biomedical-research AI for drug-discovery teams. $60M Series…

Founded 2017; GTM data orchestration — combine 100+ data sources, enrich, AI-draft outbound. $46M Series B Jan-2024 (Sequoia + Boldstart + Box Group); reported $40M+ ARR mid-2024; viral with growth…

Founded 2023; autonomous AI SRE that investigates Kubernetes + cloud-native alerts end-to-end. $10M seed Apr-2024 (Greylock + Innovation Endeavors). Direct competitor to Resolve.ai in the AI-SRE wave.

Coalfire is the largest specialist-cybersecurity-services firm in the FedRAMP/PCI/HITRUST space. Apax Partners acquired majority stake 2020. Launched AI-augmented assessment services 2024 — pentest…

Founded 2020; community intelligence platform that ingests GitHub, Slack, Discord, X, LinkedIn signals + maps them to people/companies; AI agent surfaces sales-ready signals. $32M Series B Mar-2024…

Founded 2020 (YC W21); AI-driven retail algorithmic-trading — composable strategies + LLM strategy builder. $14M Series A May-2022 (Greycroft + Y Combinator). Composer GPT + strategy-builder AI lau…

$25M Series A Jan-2025 (Menlo Ventures + Tola Capital + Pelion); founded by Monzy Merza (ex-Splunk CSO). 'Knowledge engine' that does autonomous investigation across SIEM data. Pitched as the AI an…

Datadog's AI-assistant for observability — incident summarisation, root-cause hypotheses, runbook automation. Bits AI launched Aug-2023, GA 2024 across the Datadog suite. Datadog parent: $2.7B+ rev…

Founded 2006; ABM platform with intent + AI. $175M growth round Apr-2021 ($1B val; Vista Equity). Demandbase One Smarter GTM AI launched 2024 — buyer-stage scoring + auto-personalised outbound.

DocuSign's IAM platform (launched 2024) adds AI extraction, clause analysis, and workflow automation atop the eSignature core. Acquired Lexion Jun-2024 ($165M) to consolidate the AI-CLM stack. Docu…

Founded 2021 by James Ding (ex-Allen & Overy); drafting + benchmarking AI agent that lets transactional lawyers see historical deals. $20M Series A Apr-2024 (Index Ventures + Y Combinator); ex-A&O …

Founded 2020; compliance automation (SOC 2, HIPAA, GDPR, ISO 27001) with hundreds of pre-built control libraries. $200M Series C Mar-2024 ($2B val; ICONIQ + GGV + Cowboy). Trust Hub launched 2024; …

Autonomous SOC analyst agent — investigates Tier-1 SIEM/EDR alerts end-to-end. $16.85M Series A Aug-2024 (Theory Ventures + Decibel). Founded by Edward Wu (ex-ExtraHop). Targets the 60%-of-alerts-g…

Founded 2020 by Abi Noda; engineering-productivity intelligence (DORA + SPACE metrics) with AI-driven analysis. $14M Series A Apr-2024 (Notable Capital). Differentiates with research-rigour metrics…

Dynatrace launched Davis CoPilot Sep-2023 — GenAI conversational interface on top of the (deterministic) Davis AI engine, the largest causal-AI observability deployment. Dynatrace parent: NYSE:DT, …

Founded 2021 by Dimitri Stiliadis + Varun Badhwar (ex-Palo Alto Networks Prisma Cloud); reachability-aware SCA with AI for triage. $70M Series A Oct-2023 ($200M val; Lightspeed + Coatue + Dell Tech…

OSS (Apache 2.0) LLM vulnerability scanner — 'nmap for LLMs.' Originally Leon Derczynski (Univ Sheffield / ITU Copenhagen); NVIDIA now stewards the project. Tests for jailbreaks, prompt injection, …

London-based legal AI for non-lawyer founders — templates, drafting, AI review. Acquired by Robin AI Sep-2024 (Robin = legal AI, separate funded entity). Distinct from Hugo (research) and not to be…

GitHub's CodeQL-powered auto-fix for code-scanning alerts; combines static analysis with Copilot model suggestions to generate patches for vulnerabilities. Launched Mar-2024 (public beta), GA Aug-2…

Founded 2015; the largest conversation-intelligence/revenue-intelligence platform. $7.25B val 2021 ($583M raised); Gong AI (Smart Tracker, AI Forecast, Gong Engage) 2024. Considered category-defining.

HackerOne's AI-augmented offensive security service — Hai is an LLM-assisted triage/research copilot for hackers, plus a productised 'AI red-team' service connecting customers to skilled human red-…

Founded 2020 by George Sivulka (ex-Stanford); Matrix is its enterprise research agent — multi-step research across documents. $130M Series B Jul-2024 ($700M val; Andreessen Horowitz + Index + Googl…

Founded 2019 by Snehal Antani (ex-CTO US SOCOM J6) + ex-NSA team. NodeZero is autonomous pentest — finds + exploits + verifies exploitability. $100M Series D Mar-2024 (NEA; $750M val); >300% YoY re…

HubSpot Breeze (launched Sep-2024) is the AI brand spanning Breeze Copilot (assistant) + Breeze Agents (autonomous content/prospecting/social/customer agents). HubSpot parent: NYSE:HUBS, $2.6B+ rev…

Estonian-founded legal AI startup; legal research and case analysis with EU multilingual focus. Bootstrapped; small angel rounds. Note: distinct from Hugo Health and from the unrelated 'huggo.legal…

Founded 2018; compliance-operations platform — pre-built frameworks + workflow automation. $40M Series B Dec-2022 (Riverwood Capital + Toba + Madrona). AI agent layer added 2024. Targets enterprise…

Standalone trust-center + AI-questionnaire-fill product from Hyperproof. Listed separately because the trust-center category is a distinct sub-market (Vanta, Drata, Hyperproof, SafeBase, ConveyorAI…

London-founded incident-management platform — Slack-native, fast-growing. $62M Series B Sep-2024 (Insight Partners + Index; $400M val). AI features (auto-summary, auto-postmortem, suggested-runbook…

Oslo-founded 2015; scientific research engine + AI agent for R&D. Bootstrapped + small angel rounds; broader European AI scene. Differentiates with explainable / hallucination-free claims.

Ironclad's AI agents inside its CLM (contract lifecycle management) platform. Ironclad parent: $200M+ ARR; backed by Accel, Sequoia, YC, BOND. AI features include auto-redline, review, and contract…

Tel-Aviv-founded LLM security platform — adversarial testing, prompt-injection defence, data-leakage prevention. $6M seed Nov-2023 (Entrée Capital + Samsung Next).

Founded 2020; AI email coach for sellers — Gmail/Outlook plugin scores and rewrites cold outbound. $13.2M Series A Sep-2022 (Norwest Venture Partners + LIONS Capital). $10M+ ARR reported.

Founded 2017 by Tom Martin (lawyer + technologist); LawDroid Copilot serves solo and small-firm lawyers — drafting, research, client intake automation. Bootstrapped + small angel rounds. Marketed a…

Stockholm-founded (2023) legal agent platform — drafting, review, research. $35M Series B Nov-2024 (Iconiq + Redpoint + Y Combinator; $675M val). Targets EU firms with data residency; pitched as 'E…

Seoul-founded Liner — started as web-highlighter (2015), pivoted to AI research assistant 2023. $40M Series B Jul-2023 ($350M val; Salesforce Ventures + DSC Investment + Beenext). 10M+ users.

Founded 2023; AI-agent platform for financial analysts — multi-step research, data cleansing, model-building. Small seed round; competes in the AlphaSense / Hebbia / Rogo broader space at lower pri…

Litera (Hg-backed; 50%+ of AmLaw 200 customers) added a Litera AI overlay 2024 — drafting, redlining, document analytics. Litera parent has been a private-equity rollup of legal-tech (Microsystems,…

Founded 2015; GRC platform (Risk Cloud) — flexible policy engine + workflow + risk register. $113M Series C Dec-2021 (PSG + Silversmith). AI features (auto-policy mapping, issue triage) added 2024.

Lancaster-University spinout; continuous automated AI red-teaming SaaS. $8M seed/A Jun-2024 (.406 Ventures + Atlantic Bridge). Founded by Prof. Peter Garraghan.

New Relic's AI agent — natural-language NRQL queries, error summaries, error-grouping. Launched late-2023. New Relic parent: taken private by Francisco Partners + TPG for $6.5B Nov-2023.

Numerai is a crowdsourced AI hedge fund — data scientists submit predictive signals + are paid in NMR (cryptocurrency) based on signal performance. Numerai Signals is the global-equities tournament…

Founded 2010; GRC + business-process automation. Adds AI agent (Sophia) 2024 — natural-language workflow building + automated control assessment. Bootstrapped + small minority growth equity.

Cross-listing: OpenAI Deep Research, launched Feb-2025 inside ChatGPT, is OpenAI's flagship long-horizon research agent (o3 base model). Distinct from the OpenAI Codex / Computer Use rows because t…

Founded 2021 by Didier Lopes; OSS investment-research platform — Bloomberg-Terminal-alternative. AGPLv3 + commercial OpenBB Pro. $11M seed 2022 (OSS Capital + Y Combinator). AI agent (OpenBB Copilo…

Founded 2014; AIOps platform — hybrid + multi-cloud observability with AI event correlation. **Acquired by HPE Jul-2023** (terms undisclosed; reported ~$200M+). AI agent features now integrated int…

Outreach (the largest sales-engagement platform; $4.4B val 2021) has added AI features: Smart Email Assist, Smart Account Plan, Smart Forecasting, Kaia conversation intelligence. Strategic pivot to…

PagerDuty's AIOps tier adds AI alert grouping, noise suppression, and AI agent suggestions for incident response. Includes the Jeli acquisition (May-2023, $73M) for incident retrospective AI. Pager…

Founded 2015 (ex-Pcysys); automated security validation — pentest, red-team, attack-surface validation. $60M Series D Jan-2024 ($1B val unicorn; Evolution Equity + K1). Israeli cyber pedigree; more…

Founded 2016 by Oleg Rogynskyy; revenue intelligence — activity capture + AI account intelligence. $100M Series D Sep-2021 ($1.1B val; Akkadian + ICONIQ). SalesAI agent launched 2024.

Perplexity's autonomous-research mode, launched Feb-2025 — multi-step research agent inside Perplexity. Perplexity parent: $9B val Dec-2024 ($500M round; IVP + NEA + Andreessen Horowitz).

Tel-Aviv-founded (Dor Sarig + Ziv Karliner, both ex-IDF cyber); end-to-end LLM security platform — discovery, red-teaming, runtime guardrails. $9M seed Apr-2024 (Shield Capital + Greenfield).

YC-W24 contract-review AI; auto-redlines vendor contracts based on the in-house playbook. $3.6M seed 2024 (Y Combinator + Liquid 2). Targets in-house legal at SaaS companies.

Founded 2024; AI personal-finance assistant — analyses spending, suggests savings, manages budgets via chat. Consumer-tier; small seed funding. Distinct from Pluto Money (different product; same sp…

OSS (MIT) prompt-engineering eval and red-team CLI; supports OWASP LLM Top-10 vulnerability tests, jailbreak generation, and CI/CD pipeline integration. $5M seed Mar-2024 (Andreessen Horowitz). Fou…

Founded 2022 by Ian Swanson + Daryan Dehghanpisheh (ex-AWS); AI/ML supply-chain + GenAI security. **Acquired by Palo Alto Networks Apr-2025 for $700M.** Products: Recon (red-team), ModelScan (model…

Founded 2023; AI sales-agent product — autonomous outbound, meeting booking, follow-up. $4.7M seed Jul-2024 (Lightspeed + Mantis). Direct competitor in the 11x.ai / Artisan AI / Bland niche.

Founded 2024 by Spiros Xanthos (ex-Splunk Observability SVP; founder of Omnition acq Splunk 2019). Autonomous AI SRE — investigates incidents end-to-end. $35M seed Aug-2024 (Greylock + Unusual). Pi…

Founded 2019 by Yaron Singer (Harvard); ML risk management — model testing, validation, runtime detection. **Acquired by Cisco Sep-2024** to anchor Cisco's AI security stack. Note: Harrison Chase (…

Founded 2022 by Gabriel Stengel (ex-Lazard + Stanford); AI agent purpose-built for investment-banking workflows — comps, precedent transactions, IC memo drafting. $50M Series B Mar-2025 ($350M val;…

OSS-first AI-SRE platform; CodeCollection (the OSS runbook library) is the differentiator — 'GitHub for SRE runbooks.' Apache-2.0. Bootstrapped + small angel funding. RunWhen Local self-hosted.

Tokyo-based Sakana AI's autonomous research agent — generates ideas, runs experiments, writes papers, peer-reviews. v1 paper Aug-2024 (arXiv 2408.06292); v2 update 2025. **Sakana parent: $214M Seri…

Cross-listing: Salesforce Agentforce already exists under Vertical / Customer-support / voice memory; this row is the Agent-harness framing — Agentforce as the consumer-facing agent IDE / builder o…

Founded 2018 by Josh Nicholson (PhD); citation-supported research search — find supporting/contrasting/mentioning citations for a scientific claim. **Acquired by Research Solutions Sep-2023 for $26…

Founded 2020 by Shrav Mehta; the third major compliance-automation platform (with Vanta + Drata). $56M Series B Mar-2023 (Accel + Kleiner Perkins + Base10). Comply AI launched 2024 — control gap an…

Semgrep's AI layer (Assistant) on top of its OSS rules-based SAST engine — triages findings, suppresses false positives, drafts fixes via PR. Semgrep parent: $59.5M Series C Nov-2022 ($530M val); >…

Founded 2021 inside Booz Allen Hamilton; spun out 2022. Detection-as-code platform — threat emulation + auto-generated detections. Adds GenAI threat-content layer 2024. Acquired by Cisco's Splunk M…

Snyk's AI-assisted code-fix product built on the 2020-acquired DeepCode AI engine. Combines symbolic analysis + multiple AI models for fix generation. Snyk parent: $530M raised; $7.4B val Dec-2022;…

Maker of OSS BloodHound (the Active-Directory graph tool); BloodHound Enterprise is the productised attack-path management platform. $33M Series A Aug-2024 (Insight Partners). Adds AI for attack-pa…

Stanford NLP's STORM (Synthesis of Topic Outline through Retrieval and Multi-perspective question-asking) — generates Wikipedia-style articles from a topic prompt. NAACL 2024 paper (arXiv 2402.1420…

Sysdig's GenAI-powered cloud-security agent — runtime workload protection + CNAPP findings explained. Launched May-2024. Sysdig parent: $350M Series G 2022 ($2.5B val); Permira-led 2023 round.

Dublin-founded 2018; no-code automation initially for SOC analysts. $50M Series B Dec-2023 (Felicis + Accel + CrowdStrike's Falcon Fund). AI Agent mode added 2024; Tines Workbench gives analysts an…

Tonkean is a workflow-orchestration platform; LegalWorks is the legal-ops vertical with AI agents for intake triage, contract routing, and SLA management. Tonkean: $50M Series C Sep-2022 (Insight +…

Founded 2021; PLG-revenue intelligence — predicts which free users will convert, sales-ready signals. $17.5M Series A May-2023 (Tiger Global + Sequoia + Together Fund). Pivoted to AI-agent position…

Torq's HyperSOC is the AI-augmented offering of its hyperautomation platform (no-code security workflows + LLM agents). Torq parent: $70M Series C Apr-2024 (Evolution + Bessemer + Notable; >$300M t…

Holly AI is the trading-idea generation agent inside the Trade Ideas platform (founded 2003). Holly back-tests millions of strategies overnight + surfaces 'tradeable ideas' for retail/prosumer trad…

Founded 2021 by Sravish Sridhar (ex-Kinvey founder); GRC automation + AI trust center. $25M Series A Jun-2023 (Felicis + Cowboy + Cota Capital). Differentiates with cyber-insurance + 'trust center'…

Founded 2018; the largest compliance-automation platform by valuation/headline. $150M Series C Jul-2024 ($2.45B val; Sequoia + Atlassian Ventures + CrowdStrike Falcon Fund). Vanta AI launched 2024 …

Sequoia-led $20M seed (Jun-2024); founded by Oege de Moor (Semmle / GitHub CodeQL founder). Autonomous offensive-security agent that finds and exploits web vulns end-to-end. Topped HackerOne leader…

You.com's research-agent mode — multi-step web research + synthesis. You.com parent: $99M Series B Aug-2024 ($700M val; Georgian + NEA + Salesforce Ventures). Founded by Richard Socher (ex-Salesfor…

Deep Memory feature optimises embedding space per use-case (+22% retrieval accuracy). Deep Lake PG unifies serverless Postgres (agent short-term state) + billion-scale vector search (long-term memo…

Limited native hybrid (users build RRF custom). Fast Rust core (v2.5).

Storage-optimised index scaling to billions of vectors. Hosted MCP servers for UC functions, Genie, Vector Search. Agent framework integrates retrieval with Unity Catalog governance.

Mature hybrid (BM25 + vector) implementations. RRF + weighted combination. Elastic ELSER sparse model.

Embedded vector DB (Arrow columnar). RRF reranker. Petabyte-scale on disk.

End-to-end multimodal (text + image) vector search via single API. Handles embedding generation internally. Ecommerce personalization with proprietary models.

Multi-vector columns (10 simultaneous). Native hybrid search (v2.5). CAGRA + Vamana GPU/CPU (v2.6).

Agent memory store for both short-term (document) and long-term (vector). LangGraph checkpointer for stateful agents. Vector search extended to Community Edition (Sept 2025).

Stores embeddings alongside relational + full-text data. HNSW + IVFFlat ANN indexes. Used as agent conversation memory via LangChain + MCP. Foundation of Supabase AI and many self-hosted RAG stacks.

Managed vector DB. Cascading sparse + dense + rerank pipeline; pinecone-rerank-v0 .

Distribution-Based Score Fusion + RRF. Sparse vectors native; filtering via ANN graph modification.

Hybrid vector + keyword + semantic reranking. Cortex Agents (GA Nov 2025) orchestrate over Cortex Search for unstructured data. No separate vector infra needed inside Snowflake.

Object-storage-first vector DB. Powers Notion and Cursor. SPFresh indexing; ~10× cost reduction vs. memory-resident peers.

Combines vector ANN, lexical, structured filtering, and ML-learned ranking in one distributed system. Integrated chunking + layered ranking for RAG. No separate embedding infra needed.

Hybrid Search 2.0 with learned ranking. BM25 + vector in unified index.

Clinician-assist ambient documentation. Source mapping: every AI-generated summary element traced back to the source utterance. Audit-and-trust layer over episodic memory. Built on proprietary 1.5M…

Treats memory as first-class architecture. Captures the digital footprint of every interaction; retrieves preference and history at engagement time. Public example: airline knowing a frequent flyer…

Target identification / drug repurposing / mechanism tracing. 85+ data sources, petabyte-scale, rebuilt every few weeks. Wet-lab results re-enter the graph and shift downstream predictions — instit…

Drug discovery / target identification / causal mechanism tracing. The graph is the memory: 7 years of curated biomedical cause-effect relationships compounding with each new ingestion. Scientific …

Chat Memories (user-defined facts), auto-memories for c.ai+ subscribers, pinned memories, in-context retention. PipSqueak 2 model (April 2026) reduces in-conversation drift. Memory Visualization me…

No-code interactive narrative platform. Characters track what they know, when they were lied to, what they misremember; story state changes downstream behaviour. Characters can deliberately lie / m…

Activates a Socratic interaction style — guiding questions, scaffolded concepts, knowledge checks rather than direct answers. With ChatGPT Memory enabled, persists learning goals and facts across s…

SMB law firms. AI embedded in practice management; pulls full matter overviews — WIP, billing, documents, communications — as context. Memory is the PMS data; no separately persisted cross-matter l…

Aug 2025 relaunch added agentic deep-research workflows (plan-execute-cite). Holds context within a session; no persistent cross-session matter memory layer publicly documented yet. Thomson is buil…

Enterprise contact centre (voice + chat). Short-term covers current session; long-term persists across sessions and is shared by every agent a customer might encounter. Most explicitly engineered m…

Mimir hierarchical memory: short-term (verbatim), medium-term (summaries + emotion), long-term (cross-session). Scene awareness layer adds in-engine perception. Unity + Unreal SDKs with real-time v…

Sales + contact centre. Customer history feeds real-time guidance and coaching. Memory survives handoffs between AI agent, human agent, and channels — but leans on CRM as the primary store.

Autonomous support agents resolving end-to-end (refunds, cancellations). Channel-agnostic memory: chat → call carries prior context. Agents recognise returning customers and recall prior issues.

Specialty-care ambient scribe (oncology focus). Pulls forward prior notes with precise dates; synthesises interval history (what changed since last visit); auto-populates disease-status and treatme…

Adds GPT-4-powered Roleplay (character conversations, with cross-call memory for "Lily") and Explain My Answer on top of Duolingo's existing spaced-repetition engine, which tracks per-word and per-…

Systematic literature review + data extraction + meta-analysis. Every extracted claim anchored to its exact source sentence; Notebook persists paper collections across projects. Research Agents (De…

Persistent pronunciation profile per user, tracking which sounds, words, and patterns the learner consistently mispronounces. Self-evolving model re-evaluates and updates the profile as the learner…

Plaintiff litigation. Shared Firm Knowledge (firm-wide institutional store) + autonomous Auditor agent that runs nightly across all open matters surfacing missed value drivers (TBIs, unfiled MRI or…

System 2 (7B VLM) feeds tokens to System 1 (80M visuomotor policy at 200 Hz). Temporal memory module maintains visual-state history; force-feedback buffered as tactile memory proxy. 500 hours of hu…

Embedded in case-management platform. Holds full scope of every matter — deadlines, billing, discovery, communications, negotiation history — as live context for every query. Case data is the conte…

Full autonomous-scientist loop. Owl agent surfaces prior work so downstream agents don't re-derive known results. WikiCrow produced cited factsheets for 15,616 previously undocumented human genes; …

Encodes scene into 3D Gaussian Splat representations via 3D VAE; latent Diffusion Transformer predicts future Gaussian Splat states conditioned on robot actions. 3D scene representation persists ac…

Personal lawyer memory + matter memory + institutional firm memory + client-institution memory. Opt-in at user level with admin kill-switches and ethical-wall controls. Announced Jan 2026.

Corporate / transactional due diligence. Parallel AI agents across large document sets with firm-specific configurations and review workflows. Configuration amounts to a shallow institutional memor…

Ambient documentation with clinician-profile learning. System learns each clinician's preferred note structure, phrasing patterns, symptom-outlining conventions — applies persistently across sessio…

Patient-conversation / chronic-care staffing model. HIPAA-compliant Memory Store holds non-EHR salient facts from prior calls, persists across engagements, explicitly not used for model training. S…

CX support. Currently CRM-backed customer attributes + conversation history lookup. Lifecycle-persistent memory ("priorities learned day one considered in year 10") on roadmap as of Pioneer 2025 — …

Leading purpose-built NPC engine. Long-Term Memory (named feature) stores interaction facts across sessions; surfaces them contextually. Character Brain coordinates multiple ML models. Personality …

Enterprise R&D document processing. Researcher Workspace lets users accumulate a corpus + apply self-written context filters; extracted data from experiments across hundreds of papers persists in a…

Primary care / virtual physician intake. Two-agent pipeline: one extracts the relevant EMR subset, second answers grounded to that filtered history. Architecture explicitly diverges by patient cond…

Stores full chat transcripts per student (visible to learner and teacher via dashboard). Layers over Khan Academy's existing mastery graph (exercise completion, skill progress) for cross-session to…

Therapeutic-antibody discovery in a closed Design-Build-Test-Learn loop. EVA accumulates structured outcome data from each robotic experiment cycle and feeds it back to generative ML models directi…

AI Dungeon's maker. Voyage (April 2026) launched as a creator platform for AI RPGs. World Engine (5 years in dev) is dedicated AI tracking characters, objects, backstories, relationships independen…

Architecturally notable — a named, documented service. Shared key-value store any agent (bot or human) can read/write across a customer's conversational journey, reducing repetition and enabling co…

Jan 2026 platform relaunch. Persists negotiation history, prior decisions, accepted risk positions across all contracts; new terms align with what the org has already agreed. Branded as the antidot…

Teacher productivity platform (80+ teacher tools, 50+ student tools) for planning, differentiation, IEP generation, feedback. Does not maintain a cross-session learner model. Teachers input student…

Two-tier memory bank on a VLA backbone. Perceptual MB holds low-level visual details; Cognitive MB holds high-level semantic summaries; working-memory module retrieves and fuses both into observati…

Formalises spatial localisation question answering (SLQA). High-density memory over 3D scene graphs (semantic entities + metric coordinates); LLM agent performs joint retrieval over both modalities…

Ambient scribe + longitudinal clinical intelligence (strategic partnership Jul 2025). Live encounter transcript reconciled against historical labs/imaging/notes; condition-suspecting + care-gap eng…

AI companion built around long-term memory as primary differentiator. Mind Map 2.0 gives users a navigable graph of what their Nomi knows about them and how memories connect — the most transparent …

Dual-system (VLM "System 2" + Diffusion Transformer "System 1") jointly trained end-to-end; planner outputs persist as conditioning context across action generation. N1.7 adopts relative end-effect…

Builds long-horizon memory by captioning video segments with VILA, storing captions with timestamps + 3D position coordinates in MilvusDB. At query time, LLM iterates retrieval across text, time, a…

Clinical decision support. Visits (Aug 2025) connects evidence layer to the live encounter — patient-specific context binding distinct from pure literature RAG. DeepConsult agent autonomously cross…

Enterprise voice agents. Context heavily runtime-injected from CRM / telephony / API / user history; passes context to humans on escalation. Strong on plumbing, lighter on autonomous memory accumul…

Tracks per-card answer history within a study set; computes a Memory Score for SRS scheduling. Magic Notes converts uploaded content (notes, PDFs, URLs) into flashcard sets automatically. Q-Chat, t…

Voice automation (payment, scheduling, order status). Memory primarily within-session with structured handoff notes; the 500M+-conversation corpus is training data, not runtime memory.

Decade-old AI companion. Memory bank logs personal facts, preferences, significant moments; writes a Diary after sessions. Memory reliability documented as weak (64% user dissatisfaction 2025) but …

Literature discovery + citation mapping. Oct 2025 revamp formalises checkpointed search paths as named "rabbit holes" — each checkpoint is a saved state of the citation graph the user can branch fr…

Agentforce agents grounded in Data Cloud — unified customer record across silos. Memory is the CRM record made available to AI; not an independent memory system. AI reads/writes through the existin…

Mirrors human cognitive functions: explicit memory, vision, hearing, tactile. Translates natural-language instructions into actions with explainable intermediate reasoning. Memory stores task proce…

Teacher-created "Spaces" (structured AI workspaces) where Dot, the student-facing assistant, guides learners through agendas. Student interaction data — completion, errors, time-on-task — surfaces …

Stores surrounding textual context of every citation plus a deep-learning classification (supports / disputes / mentions). Persisted across the database, not computed on demand. MCP integration (20…

Incrementally constructs hierarchical 3D scene graph during navigation — nodes for objects/groups/rooms, edges for spatial relations. Prunes stale nodes with linear complexity. Live graph passed as…

Bret Taylor + Clay Bavor. Agent Data Platform links unstructured conversation data to enterprise systems (billing, inventory, policies, transactions); persists across channels and time. Memory accu…

Real-time simultaneous inference + data collection — dynamically updates in-context representation as the robot encounters new situations. Lets the system adapt to novel embodiments / unseen enviro…

NZ/US digital-humans company. Patented Digital Brain simulates emotional state and learns from interactions. Cross-session Memories layer for user- and persona-specific context persistence. Deploye…

Onboarding assessment builds an initial learner profile and generates a personalized curriculum; AI supports open-ended conversational practice with real-time speech recognition. No SRS or long-ter…

Transactional / contract drafting. Ingests entire precedent corpus (OneDrive, Dropbox); ML surfaces and reuses the firm's own prior clauses during live drafting in Word, tuned to specific deal cont…

Authoring toolkit + SDK for narrative designers. Tracks what each character knows, doesn't know, has witnessed; explicit handling of NPC knowledge inconsistency across sessions. Used in AR/VR and A…

Doctor-assist ambient documentation. Pulls forward prior documentation into a pre-visit synthesis rather than waiting for the encounter. Patient summaries surface relevant insight from history befo…

Embeds micro-assessments into every lesson to track understanding in real time, adjusting difficulty and teaching approach within session. Combines AI personalization with human-curated content. No…

Single end-to-end neural network ingesting multi-camera video + navigation + ego-motion. Maintains evolving 3D occupancy volume (voxel-level) + 3D Gaussian rendering as internal world state — spati…

All surround-camera views encoded through video tokeniser to continuous latent; past latent sequences serve as explicit temporal context via space-time factorised transformer. GAIA-3 (15B params, 2…

Extends π0 with MEM layer giving the policy short-term (within-task) + long-term (cross-task, >10-min horizon) memory. Co-trained on robot teleoperation + human video + text. Hierarchical inference…

Voice-AI platform with custom STT + LLM + TTS stack ('the whole pipeline'). Raised $40M Series A Apr-2024 (Lightspeed + Anthology). Distinguished by 'pathway' graphs (deterministic call flows) and …

CallRail's AI-driven conversation intelligence layer atop their call-tracking product. Marketing-vertical voice analytics — auto-tagging, sentiment, lead scoring. CallRail is profitable / private (…

Cartesia's Sonic TTS — state-space-model-based voice generation, lowest published TTS latency (~40ms). Cartesia raised $64M Series A Q1 2025 (Index + Lightspeed); founded by Mamba authors. Agent pl…

ElevenLabs launched Conversational AI Nov-2024 — STT + LLM + their own market-leading TTS in one pipeline. Combines vendor's voice cloning + multilingual TTS strengths with hosted call orchestratio…

Hume's Empathic Voice Interface (EVI) — voice agent platform with vocal-affect understanding. Hume raised $50M Series B Mar-2024 (EQT). Distinguished from pure LLM voice agents by emotion/prosody m…

LiveKit Agents — agent framework built on LiveKit's open-source WebRTC media server. LiveKit raised $45M Series B Sep-2024 ($300M val). Powers OpenAI Realtime API's WebRTC transport (LiveKit is the…

OpenAI's Realtime API for native speech-to-speech multimodal interaction with GPT-4o — launched Oct 2024 (DevDay). Skips the STT → LLM → TTS pipeline by passing speech directly to GPT-4o. The platf…

Open-source Python framework for real-time voice + multimodal AI agents, maintained by Daily.co. Pipeline-style architecture (STT → context → LLM → TTS) wired over WebRTC. The OSS substrate that ma…

Voice-agent platform with vendor focus on sub-800ms turn latency and proprietary voice models. Y Combinator S24; raised $4.6M seed Aug 2024 (BoxGroup + others); >$1M ARR by month 6 per founders. Di…

Speechmatics (UK; STT leader) launched Flow Q4 2024 — conversational voice-agent platform composed on Speechmatics STT + caller LLM. Speechmatics raised $62M Series B (2022); enterprise + media cus…

No-code voice-AI platform — drag-and-drop call-flow builder targeted at non-developers. Berlin-based; raised $7.4M Series A Sep-2024 (Singular). Distinguishes itself from Vapi/Retell with visual ca…

Voice-AI developer platform — STT + LLM + TTS pipelines wired to telephony (Twilio / SIP). Used to build outbound + inbound phone agents at scale. Raised $20M Series A from Bessemer + Y Combinator …

Vonage's AI Studio — visual conversational-agent designer that runs on Vonage CPaaS (SMS / Voice / WhatsApp). Vonage acquired by Ericsson 2022 ($6.2B); AI Studio is the conversational AI layer of t…

Small wearable (pendant / clip / bracelet). Continuous capture; converts speech to text immediately; never stores raw audio . Daily summaries, personal fact sheet, to-do suggestions. Cloud processi…

Software platform / OS layer for AI wearables (glasses, rings, pendants, speakers) that abstracts hardware from AI orchestration. Memory and model providers are pluggable and user-controlled rather…

Pendant necklace (~2 inches). Always-on listening feeds a persistent AI companion persona. Companionship-framed rather than productivity / recall. Cloud LLM backend; no detailed encryption claims.

Mac/iOS app (no hardware). Captures meeting audio via system-audio tap (no bot joining); generates notes from your shorthand + audio. Local audio processing claim; cloud for summaries. Enterprise t…

Continuously captures screen + audio + wearable input; AI search across the timeline. Privacy-first (local-first storage). Originated as Rewind macOS app; pivoted to wearable + lifelog product.

Developer-first reference platform for AI wearables. LVMM 2.0 captures visual data through video, encodes it into frames on-device, and stores moments as searchable memories; companion app surfaces…

Wearable orb (necklace or clip) with open-source hardware + firmware. Continuous audio capture; real-time GPT-4o transcription; searchable memory graph. Supports app screen capture too. Hybrid priv…

Joins Zoom / Meet / Teams as a meeting bot. Real-time transcription, speaker identification, searchable archive across past meetings. SSO + admin controls at Business tier.

Magnetic clip-on wearable. Records in-person conversations and calls; AI summaries, action items, searchable transcripts. AES-256 + TLS 1.2 in transit; doesn't retain audio after transcription. Als…

Meeting bot + browser extension. Captures audio + video; engagement scoring (talk-time, sentiment, attention); enterprise search across all past meetings. "Ada" digital-twin feature builds an AI me…

Smart ring with voice-activated capture (tap to record, not always-on). Personal knowledge graph with companies, people, meetings; pulls context from past meetings and shared interests. Ex-Meta eng…