Open-source agent memory systems

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

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

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.

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

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

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

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

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

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).

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). EU-headquartered alternative to US/Ch…

"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.

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

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

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

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

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 cross-embodiment robot trajectories. Rais…

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.

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

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

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 variant for premium tier. No frontier-tier op…

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

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 MoE, Qwen 3 235B-A22B MoE. Qwen-Coder, Q…

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

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 (Felicis); cloud product at cloud.browser-…

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-flight to extend session length. Captures wh…

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

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: Project/Package/File/Module/Class/Function/Met…

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), Command R+ (104B MoE), Command R7B (smalle…

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.

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

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 moment' in markets Jan-27-2025 — NVIDIA d…

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

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

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 code-agent products including Anthropic…

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 RAM at near-zero runtime overhead.

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

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 "fact true."

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

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

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 (HopeJR, SO-100 arms). LeRobot is the de-f…

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

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

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 .

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 substrate OpenAI chose). Agents framewo…

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

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

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.

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

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 CLI sessions can be "teleported" to the…

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

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 privacy: cloud by default (SOC 2/HIPAA/AES-2…

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

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-series and Codex (model) for execution.…

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 Hands AI cloud (managed SaaS).

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

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

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 many of the voice platforms build on or co…

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

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

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

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-up button')). Used as the OSS substrate …

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

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 advantage. Not a startup — division of …

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.

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

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 markdown + Threads form the memory layer…

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).

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 the user's screen. SOTA on OSWorld at rel…

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

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

Meta-programming for a multi-agent collaborative framework.

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 2.0 (Yi-34B and base sizes). Important …

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

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-management layer — task context, instruction sta…

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

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 snapshots. .aider.conf.yml carries stati…

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 CONVENTIONS.md and .aider.conf.yml — author-edit…

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