| π0.5 (Physical Intelligence) | T1 | Multi-Scale Embodied Memory (short + long episodic) ↗ 2 | vector | attentionsimilarity | long-term | append-onlyconsolidation | episodetrajectory | opaque | append | 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: VLA predicts subtasks as text tokens, then executes as action chunks. ↗ 2 | Open-world generalisation to unseen household environments. Cleans an entirely new kitchen with no environment-specific training. ↗ 2 | vision + sensor + text ↗ 2 | 2024 (Physical Intelligence formally founded 2024; informal work began Q3 2023; Chelsea Finn Sergey Levine Karol Ha… ↗ 2 | not applicable — not OSS 3 | not applicable — not OSS 3 | not applicable — no GitHub repo 3 | searched not found ↗ 2 | not applicable — not academic 3 | $400M total $2.0B val Series A · 2024-11 ↗ 2 | 51-200 ppl ↗ 2 | searched not found ↗ 2 | searched not found ↗ 2 | Trained on Open X-Embodiment + proprietary in-house datasets; open-source weights available ↗ ● | Cloud inference + on-robot deployment (websocket streaming) ↗ ● | US (San Francisco) ↗ 2 | Physical Intelligence; founders Sergey Levine (UC Berkeley robotics professor) Karol Hausman Ch… ↗ 2 | no scale claim ↗ 2 | π0.5 (Physical Intelligence): open-world generalisation results — clean unseen kitchens (vendor paper) ↗ 2 | not applicable — not a research paper 3 | not applicable — not a research paper 3 | searched not found ↗ 2 | searched not found ↗ 2 | searched not found ↗ 2 | searched not found ↗ 2 | searched not found ↗ 2 | searched not found ↗ 2 | searched not found ↗ 2 | searched not found ↗ 2 | not applicable — embodied policy, not a database 2 | model-version semantics (policy checkpoints) ↗ 2 | not applicable — embodied policy, not a database 2 | not applicable — embedded policy weights, no schema 2 | robot-per-instance (per-robot memory) ↗ 2 | always append (no resolution) ↗ 2 | episode-based memory; long-term memory module retains landmarks; episodic memory pruned per task ↗ 2 | no MCP support advertised — vertical product, no MCP server / client integration documented ↗ 2 | no A2A protocol support advertised — vertical product, no A2A integration documented ↗ 2 | no OpenTelemetry integration advertised — vendor logs/observability not publicly documented ↗ 2 | no public webhook API advertised — vendor product, no public docs for webhook integration ↗ 2 | no public import/export API advertised — vendor product, data movement via enterprise integration only ↗ 2 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | robotics (humanoid, manipulation), autonomous-driving, embodied AI research ↗ 2 | weeks-to-months (custom robot/vehicle integration) or research-only code release ↗ 2 | not for non-embodied / non-physical use cases ↗ 2 | real-time perception + spatial reasoning + multi-modal sensor fusion ↗ 2 | requires robot / vehicle hardware + sensor stack ↗ 2 | Generalist robot policy with strong simulation-to-real transfer; memory is implicit in the policy network rather than explicit external store. ↗ 2 | Implicit memory is harder to inspect or edit; embodied agent only — not generalizable to LLM agents. ↗ 2 | π0.5 blog paper ↗ 2 |
| 01.AI Yi family stale | T2 | Open-weights frontier-tier Chinese model family (Yi-Lightning / Yi-34B) ↗ 2 | parametrickv-cache | parametric-recallattention | parametric-permanent | read-only | weightkv-token | opaque | n/a | 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 second Chinese open-weights option alongside DeepSeek + Qwen. ↗ 2 | Yi-Lightning matched GPT-4o on LMSYS Arena Oct-2024 at fraction of training cost; Yi-34B Apache 2.0; trained at $3M total compute claimed; Kai-Fu Lee founder pedigree (ex-Microsoft Research Asia, Google China) ↗ 2 | text, code; Yi-VL (vision) variant ↗ 2 | 2023-07 (01.AI founded); 2023-11 (Yi-34B); 2024-05 (Yi-Large); 2024-10 (Yi-Lightning) ↗ 2 | Yi-Lightning (2024-10) ↗ 2 | Apache 2.0 (Yi-34B + base sizes); proprietary (Yi-Large, Yi-Lightning) ↗ 2 | github.com/01-ai/Yi ~7k+ stars ↗ 2 | Top-3 Chinese model lab; LMSYS arena visibility from Yi-Lightning launch Oct-2024; meaningful but not dominant OSS HuggingFace presence ↗ 2 | not applicable — not academic 3 | $1B+ raised; >$1B valuation 2024 (Alibaba, Lightspeed China, Tencent investors) ↗ 2 | Chinese enterprise via 01.AI API; some Western OSS via HuggingFace (Yi-34B) ↗ 2 | API: Yi-Lightning $0.14 per 1M tokens (vendor-claimed lowest among frontier-tier); Yi-34B weights free under Apache 2.0 ↗ 2 | China-based; not US-certified ↗ 2 | Self-hosting recommended for Western sensitive data; Apache 2.0 weights for Yi-34B remove hosted dependency ↗ 2 | Self-hostable (Yi-34B Apache 2.0) + 01.AI API for Yi-Large / Yi-Lightning; Together AI hosts Yi-34B ↗ 2 | Beijing, China ↗ 2 | Kai-Fu Lee (CEO; ex-Microsoft Research Asia President, ex-Google China President, Sinovation Ventures) ↗ 2 | not applicable — substrate foundation model 3 | LMSYS Chatbot Arena: Yi-Lightning matched GPT-4o (Oct-2024); MMLU ~78% (Yi-Large); MMLU-Pro ~70%; HumanEval ~85% ↗ 2 | not applicable — not a research paper 3 | open weights for Yi-34B/9B/6B/Coder; proprietary for Yi-Large/Lightning ↗ 2 | REST + SDK; HuggingFace Transformers for Yi-34B ↗ 2 | not applicable — substrate foundation model 3 | not applicable — substrate foundation model 3 | not applicable — substrate foundation model 3 | not applicable — substrate foundation model 3 | not applicable — substrate foundation model 3 | not applicable — substrate foundation model 3 | not applicable — not a memory product 3 | not applicable — substrate foundation model 3 | not applicable — substrate foundation model 3 | not applicable — not a memory product 3 | not applicable — not a memory product 3 | not applicable — substrate foundation model 3 | not applicable — not a memory product 3 | not applicable — not a memory product 3 | not applicable — substrate foundation model 3 | not applicable — substrate foundation model 3 | not applicable — substrate foundation model 3 | not applicable — substrate foundation model 3 | not applicable — substrate foundation model 3 | not applicable — substrate foundation model 3 | not applicable — substrate foundation model 3 | not applicable — substrate foundation model 3 | not applicable — substrate foundation model 3 | not applicable — substrate foundation model 3 | not applicable — substrate foundation model 3 | not applicable — substrate foundation model 3 | open (Apache 2.0 for Yi-34B); single-vendor for Yi-Lightning ↗ 2 | any (HuggingFace, vLLM, llama.cpp) ↗ 2 | not applicable — substrate foundation model 3 | Chinese enterprise; cost-sensitive OSS deployers globally ↗ 2 | <10min via 01.AI API; <1hr self-host (Yi-34B Ollama / vLLM) ↗ 2 | China data-residency for regulated Western workloads; Yi-Lightning closed weights; less mindshare in West than Qwen / DeepSeek ↗ 2 | cost-efficient frontier inference; Chinese-language native; Apache 2.0 for Yi-34B tier ↗ 2 | HuggingFace; Together AI hosting; Sinovation Ventures alumni network ↗ 2 | Kai-Fu Lee founder pedigree; Yi-Lightning matched GPT-4o on LMSYS at fraction of cost; Apache 2.0 Yi-34B is broadly deployable. ↗ 2 | Less Western mindshare than DeepSeek / Qwen; Yi-Lightning closed-weights; smaller OSS ecosystem; China data-residency concerns. ↗ 2 | www.01.ai/ ↗ 2 |
| 11x.ai | T1 | AI sales agents (Alice, Mike, Jordan) ↗ 2 | none-trivial | none | session-only | agent-controlled | turn | opaque | out-of-scope | 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 AI-employee startups of 2024. ↗ 2 | $50M Series B Oct-2024 ($350M val; Benchmark + a16z); >$10M ARR Q3-2024; thousands of customers ↗ 2 | not applicable — wrong section 3 | 2022 (founded) ↗ 2 | not applicable — wrong section 3 | not applicable — not OSS 3 | not applicable — no GitHub repo 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | $50M Series B Oct-2024 ($350M val; Benchmark + a16z) ↗ 2 | Thousands of B2B SaaS customers ↗ 2 | Per-seat / per-agent enterprise tiers ↗ 2 | not applicable — wrong section 3 | not applicable — wrong section 3 | SaaS ↗ 2 | San Francisco / London ↗ 2 | Hasan Sukkar (CEO) ↗ 2 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — not OSS 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — not a memory product 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — not a memory product 3 | not applicable — not a memory product 3 | not applicable — wrong section 3 | not applicable — not a memory product 3 | not applicable — not a memory product 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | autonomous outbound SDR / RevOps ↗ 2 | not applicable — wrong section 3 | Largest AI-SDR by mindshare 2024; Benchmark + a16z; 'AI employees' framing resonates. ↗ 2 | Ongoing controversy about ICP fit and churn (TechCrunch reporting late-2024); deliverability concerns at high outbound volumes. ↗ 2 | www.11x.ai/ ↗ 2 |
| 1X Technologies | T2 | Humanoid robot maker (Neo / Eve) — OpenAI portfolio ↗ 2 | weight | parametric-recall | parametric-permanent | agent-controlled | trajectory | opaque | training-time | 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 Series A (Mar-2023, OpenAI lead). Distinguishing feature: consumer-home positioning rather than industrial. ↗ 2 | $100M Series B Jan-2024 (EQT + OpenAI); home-consumer humanoid (Neo) targeting 2025+ ↗ 2 | vision + language + bimanual action ↗ 2 | 2014 (founded as Halodi Robotics) ↗ 2 | Neo (consumer humanoid, 2025-2026 launch) ↗ 2 | not applicable — not OSS 3 | not applicable — no GitHub repo 3 | Strong viral demos; consumer-home angle differentiates from Figure/Apptronik ↗ 2 | not applicable — not academic 3 | $100M Series B Jan-2024 (EQT + OpenAI Startup Fund); $40M Series A 2023 ↗ 2 | Eve pilot deployments; Neo pre-orders / waitlist for consumer ↗ 2 | not applicable — wrong section 3 | not applicable — wrong section 3 | Internal trajectory data; closed weights ↗ 2 | Direct customer (industrial Eve + consumer Neo) ↗ 2 | Norway (Moss) + US (Sunnyvale) ↗ 2 | Bernt Børnich (CEO; founded as Halodi Robotics 2014) ↗ 2 | Eve pilots (~10s of units); Neo pre-orders ↗ 2 | Demo videos: pour, fold, manipulate; no public benchmarks ↗ 2 | not applicable — not a research paper 3 | not applicable — not OSS 3 | No public API ↗ 2 | searched not found ↗ 2 | searched not found ↗ 2 | Internal 1X cloud ↗ 2 | Per-customer fleet ↗ 2 | Enterprise-grade ↗ 2 | not applicable — wrong section 3 | not applicable — not a memory product 3 | searched not found ↗ 2 | Eve / Neo generation versioning ↗ 2 | not applicable — stateless 3 | not applicable — not a memory product 3 | per-customer fleet ↗ 2 | not applicable — stateless 3 | not applicable — stateless 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — stateless 3 | not applicable — not OSS 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | industrial (Eve) + consumer-home (Neo) ↗ 2 | Months (industrial deployment); Neo home-consumer launch not GA as of 2025 — pre-order waitlist only ↗ 2 | Consumer-home humanoid is unproven category; closed platform ↗ 2 | Consumer-home humanoid (Neo) vs Figure's industrial focus ↗ 2 | OpenAI Startup Fund portfolio; Norwegian / US robotics ecosystem ↗ 2 | OpenAI-backed; consumer-home angle is differentiated; viral demos. ↗ 2 | Smaller funding than Figure; consumer humanoid market unproven; closed weights. ↗ 2 | 1x.tech ↗ 2 |
| 6sense | T2 | ABM intent + AI agent platform ↗ 2 | none-trivial | none | session-only | agent-controlled | turn | opaque | out-of-scope | 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. ↗ 2 | $200M Series E Jan-2022 ($5.2B val); 1500+ enterprise customers; Revenue AI agents 2024 ↗ 2 | not applicable — wrong section 3 | 2013 (founded) ↗ 2 | not applicable — wrong section 3 | not applicable — not OSS 3 | not applicable — no GitHub repo 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | $200M Series E Jan-2022 ($5.2B val; Blue Owl + SoftBank + Insight) ↗ 2 | 1500+ enterprise (Cisco, Dell, Workday, Mastercard) ↗ 2 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | San Francisco ↗ 2 | Amanda Kahlow (founder); Jason Zintak (CEO) ↗ 2 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — not OSS 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — not a memory product 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — not a memory product 3 | not applicable — not a memory product 3 | not applicable — wrong section 3 | not applicable — not a memory product 3 | not applicable — not a memory product 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | ABM + intent data + AI revenue agents ↗ 2 | not applicable — wrong section 3 | Strongest ABM/intent dataset; $5.2B val; broad enterprise base; AI agents on top of high-quality data. ↗ 2 | Pricing complexity; AI agents are overlay; competition with Demandbase in same niche. ↗ 2 | 6sense.com/ ↗ 2 |
| A-MEM | T3 | Atomic-note / Zettelkasten-style ↗ ● | graph | graph-traversalsimilarity | long-term | append-only | fact | inspectable | n/a | Treats memories as atomic linkable notes — explicit nod to Zettelkasten knowledge management. Dynamic linking; retroactive memory revision. ↗ ● | Atomic-note (Zettelkasten-style) memory: dynamic linking and retroactive revision of memories. 871 stars; NeurIPS 2025; 443 cites. ↗ 2 | searched not found ↗ 2 | 2025-01 ↗ 2 | no releases ↗ ● | MIT ↗ ● | 871★ +10/mo Python ↗ ● | searched not found ↗ ● | 443 cites 443/yr ↗ 2 | not applicable — not commercial 3 | not applicable — not commercial 3 | searched not found ↗ 2 | not applicable — not commercial 3 | not applicable — not commercial 3 | searched not found ↗ ● | not applicable — not a company 3 | no founder data — research paper / preprint only (no commercial entity). ↗ ● | searched not found ↗ 2 | LongMemEval: reports 60.6% overall accuracy with GPT-4o-mini (atomic-note + Zettelkasten linking); LoCoMo ablations show consistent gains. ↗ 2 | Too recent ↗ 2 | Code public + weights closed ↗ ● | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not specified — research paper does not address this operational dimension (this is a paper/system, not a productised offering). ↗ 2 | immutable append-log ↗ ● | append-only no delete ↗ ● | schema with migration ↗ ● | not specified — research paper does not address this operational dimension (this is a paper/system, not a productised offering). ↗ 2 | always append (no resolution) ↗ ● | none (permanent) ↗ ● | not applicable — research paper, no deployed product 3 | not applicable — research paper, no deployed product 3 | not applicable — research paper, no deployed product 3 | not applicable — research paper, no deployed product 3 | not applicable — research paper, no deployed product 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | research only ↗ ● | not applicable - research paper / code release 3 | not applicable - research paper 3 | research positioning (see memory_model) ↗ ● | not applicable - research paper 3 | Adaptive memory with explicit policy for what to remember vs forget. ↗ ● | Research-only; no production deployment. ↗ ● | GitHub WujiangXu/A-mem arxiv 2502.12110 NeurIPS 2025 poster ↗ ● |
| Abridge | T1 | Grounded-transcript provenance ↗ 2 | filevector | similarity | long-term | append-only | episodedocument | auditable | editor-in-the-loop | 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+ medical-encounter dataset. ↗ 2 | Deployed at major academic health systems. Inpatient + outpatient tools launched 2025. Altais physician-burnout partnership. ↗ 2 | voice + text ↗ 2 | 2025-02 ↗ 2 | not applicable — not OSS 3 | not applicable — not OSS 3 | not applicable — no GitHub repo 3 | searched not found ↗ 2 | not applicable — not academic 3 | $616M total $5.3B val Series E Extension · 2026-04 ↗ 2 | 51-200 ppl ↗ 2 | Enterprise only ↗ 2 | SOC 2 Type II, ISO 27001, HIPAA (BAA required) ↗ 2 | Trains on de-identified patient data only (raw PHI never used for training; BAA governs all customer PHI processing) ↗ 2 | Managed-only ↗ 2 | US ↗ 2 | UPMC / CMU-backed; founder Shiv Rao (UPMC practicing cardiologist; ex-UPMC corporate VC; helped… ↗ 2 | 1.5M+ medical-encounter dataset (vendor claim) ↗ 2 | no public benchmark scores found — vendor-claim 1.5M+ medical-encounter dataset; no published WER / SOAP-accuracy metric ↗ 2 | not applicable — not a research paper 3 | not applicable — not a research paper 3 | searched not found ↗ 2 | searched not found ↗ 2 | searched not found ↗ 2 | searched not found ↗ 2 | US-based HIPAA-secure data centers; tenant-per-customer logical isolation; BAA with each enterprise customer ↗ 2 | At-rest + in-transit (HIPAA-compliant US data centers); SOC 2 Type II ↗ 2 | searched not found ↗ 2 | searched not found ↗ 2 | strong consistency required for clinical record integrity ↗ 2 | immutable audit log (HIPAA + 21 CFR Part 11 require audit trails) ↗ 2 | audit-logged soft delete (regulatory retention requirement) ↗ 2 | FHIR/HL7-aligned schema (clinical standard); minor versioning ↗ 2 | tenant-per-customer (BAA-scoped); patient-scoped within tenant ↗ 2 | always append (no resolution) ↗ 2 | regulated retention (7+ years per HIPAA + state law); patient-requested deletion governed by HIPAA right-to-access ↗ 2 | no MCP support advertised — vertical product, no MCP server / client integration documented ↗ 2 | no A2A protocol support advertised — vertical product, no A2A integration documented ↗ 2 | no OpenTelemetry integration advertised — vendor logs/observability not publicly documented ↗ 2 | no public webhook API advertised — vendor product, no public docs for webhook integration ↗ 2 | no public import/export API advertised — vendor product, data movement via enterprise integration only ↗ 2 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | healthcare (ambient documentation, clinical decision support, primary care, oncology) ↗ 2 | <1week to weeks-to-months (HIPAA-regulated enterprise procurement + EHR integration) ↗ 2 | not for non-healthcare verticals; must operate under HIPAA / regional health regulation ↗ 2 | HIPAA compliance + clinical-grade provenance + EHR integration ↗ 2 | requires EHR integration + clinical workflows ↗ 2 | Strongest published evidence for clinical-encounter memory accuracy; multi-EMR integrations and large hospital deployments. ↗ 2 | Enterprise sales motion only; longitudinal cross-visit memory layered on top of single-encounter scribing rather than the architecture's primary unit. ↗ 2 | abridge.com ↗ 2 |
| Accelerate (Hugging Face) | T1 | Distributed training abstraction ↗ ● | n/a | n/a | n/a | n/a | n/a | n/a | n/a | Hugging Face's abstraction over distributed-training backends (DDP, FSDP, DeepSpeed, Megatron). Minimal code change to scale a PyTorch script. ↗ ● | ~7.7k★; standard distributed wrapper for HF Transformers. ↗ ● | searched not found ↗ ● | 2021 ↗ ● | searched not found ↗ ● | Apache-2.0 ↗ ● | searched not found ↗ ● | searched not found ↗ ● | searched not found ↗ ● | searched not found ↗ ● | searched not found ↗ ● | searched not found ↗ ● | searched not found ↗ ● | searched not found ↗ ● | searched not found ↗ ● | searched not found ↗ ● | searched not found ↗ ● | searched not found ↗ ● | searched not found ↗ ● | searched not found ↗ ● | Active — last commit 2026-05 · 9,679★ ↗ ● | searched not found ↗ ● | searched not found ↗ ● | searched not found ↗ ● | searched not found ↗ ● | searched not found ↗ ● | searched not found ↗ ● | searched not found ↗ ● | not applicable — not a memory product 3 | not applicable — not a memory product 3 | searched not found ↗ ● | not applicable — not a memory product 3 | not applicable — not a memory product 3 | not applicable — not a memory product 3 | not applicable — not a memory product 3 | not applicable — not a memory product 3 | searched not found ↗ ● | searched not found ↗ ● | searched not found ↗ ● | searched not found ↗ ● | searched not found ↗ ● | searched not found ↗ ● | searched not found ↗ ● | searched not found ↗ ● | searched not found ↗ ● | searched not found ↗ ● | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | searched not found ↗ ● | searched not found ↗ ● | searched not found ↗ ● | searched not found ↗ ● | searched not found ↗ ● | searched not found ↗ ● | searched not found ↗ ● | searched not found ↗ ● |
| ACON | T4 | Context-compression optimisation ↗ 2 | kv-cache | attention | session | evict-oldest | kv-token | inspectable | llm-arbitrate | Optimises context compression for long-horizon LLM agents. ↗ 2 | Reduces memory usage by 26-54% (peak tokens) while preserving task performance; preserves >95% of accuracy when distilled into smaller compressors; up to 46% gain for smaller LMs as long-horizon agents. Headline: 54.5% peak-token reduction on 8-objective QA while surpassing uncompressed EM/F1; baselines: No-Compression, FIFO, Retrieval, LLMLingua, naive Prompting; primary datasets: AppWorld, OfficeBench, 8-objective QA. ↗ 2 | text ↗ 2 | 2025-10 ↗ 2 | not applicable — not OSS 3 | not applicable — not OSS 3 | not applicable — no GitHub repo 3 | 28 cites · 3.9/mo · 70* / 10 forks ↗ 2 | 27 cites 27/yr ↗ 2 | not applicable — not commercial 3 | not applicable — not commercial 3 | not applicable — not commercial 3 | not applicable — not commercial 3 | not applicable — not commercial 3 | not applicable — not a deployable product 3 | not applicable — not a company 3 | no data — searched arxiv.org/abs/2510.00615; author affiliations not found in preview ↗ 2 | no scale claim — research ↗ 2 | 54.5% 8-objective QA peak-token reduction ↗ 2 | Original team only — 10 forks of https://github.com/microsoft/acon (largely fork-only); published 2025-10-01 (7 mo ago) ↗ ● | Code public + weights closed — https://github.com/microsoft/acon · 70* / 10 forks · last commit 2025-10 · MIT ↗ ● | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | strong ↗ 2 | none ↗ 2 | hard-delete supported ↗ 2 | schema locked ↗ 2 | not applicable 3 | LLM resolves and overwrites ↗ 2 | LRU eviction ↗ 2 | not applicable — research paper, no deployed product 3 | not applicable — research paper, no deployed product 3 | not applicable — research paper, no deployed product 3 | not applicable — research paper, no deployed product 3 | not applicable — research paper, no deployed product 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable - research paper 3 | not applicable - research paper 3 | not applicable - research paper 3 | not applicable - research paper 3 | not applicable - research paper 3 | Agent-context-orchestration memory. ↗ 2 | Preprint; orchestration overhead in agent loops. ↗ 2 | arxiv 2510.00615 ↗ 2 |
| Activeloop Deep Lake | T2 | Multimodal vector + serverless Postgres ↗ 2 | vectorcolumn | similarity | long-term | overwriteappend-only | chunkdocument | inspectable | overwrite | 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 memory). ↗ 2 | +22% retrieval accuracy claim. Reports 80% cheaper than comparable vector DBs. ↗ 2 | text + image ↗ 2 | 2018 ↗ 2 | not applicable — not OSS 3 | not applicable — not OSS 3 | not applicable — no GitHub repo 3 | n/a (no top PyPI pkg) 1 press ~30 (Activeloop jobs 2 | not applicable — not academic 3 | ~$20M total (Seed + $11M Series A) ↗ 2 | searched not found ↗ 2 | Free + paid ↗ 2 | SOC 2 Type II; SAML/RBAC features ↗ 2 | searched not found ↗ 2 | Both ↗ 2 | US ↗ 2 | Davit Buniatyan (founder/CEO; Princeton origin) ↗ 2 | no scale claim ↗ 2 | Deep Lake reports +22% retrieval accuracy vs Pinecone/Weaviate on enterprise RAG workloads; 80% cheaper cost claim. ↗ 2 | not applicable — not a research paper 3 | not applicable — not a research paper 3 | REST, SDK: Python, JS/TS ↗ 2 | searched not found ↗ 2 | searched not found ↗ 2 | custom (Deep Lake columnar format on object storage) ↗ 2 | hard-isolation ↗ 2 | at-rest + in-transit ↗ 2 | SSO + RBAC ↗ 2 | multiple supported ↗ 2 | not documented in public vendor materials (operational dimension typically managed in private deployment guides). ↗ 2 | immutable append-log ↗ 2 | not documented in public vendor materials (operational dimension typically managed in private deployment guides). ↗ 2 | schema with migration ↗ 2 | per-tenant ↗ 2 | always overwrite (last-write-wins) ↗ 2 | none (permanent) ↗ 2 | no first-party MCP adapter published as of 2026-05; community connectors may exist. ↗ 2 | no Google A2A (Agent2Agent) integration documented as of 2026-05. ↗ 2 | no first-party OpenTelemetry exporter documented; standard logs/metrics typically available. ↗ 2 | no ↗ 2 | Deep Lake format, Parquet, NumPy, COCO ↗ 2 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | RAG developers, agent builders (broad horizontal); financial-services, healthcare, legal, e-commerce ↗ 2 | <10min (managed cloud) to <1hr (self-host Docker) ↗ 2 | not for relational / graph-heavy queries (vector-first by design) ↗ 2 | low-latency similarity search + scale ↗ 2 | BYO embedding model + LLM; this IS the vector store ↗ 2 | Multi-modal data lake (vectors + images + video) with version control — strong for ML data pipelines plus memory. ↗ 2 | Less developer-API-friendly than Pinecone / Weaviate; positioning straddles ML training and runtime memory. ↗ 2 | activeloop.ai ↗ 2 |
| Acuvity (now Proofpoint) | T1 | MCP / Shadow-AI runtime enforcement ↗ 2 | n/a | agentic | session | read-only | prompt | auditable | n/a | 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 infrastructure layer where memory is most exposed. ↗ 2 | Acquired by Proofpoint Feb 2026 for integration into enterprise DLP / governance. ↗ 2 | text ↗ 2 | 2023 (founded by cybersecurity veterans; acquired by Proofpoint Feb 12 2026) ↗ 2 | not applicable — not OSS 3 | not applicable — not OSS 3 | not applicable — no GitHub repo 3 | searched not found ↗ 2 | not applicable — not academic 3 | no public funding data found — pre-acquisition private company; acquired by Proofpoint for undisclosed amount ↗ 2 | searched not found ↗ 2 | enterprise pricing via Proofpoint (not public); contact sales ↗ 2 | Proofpoint enterprise compliance (SOC 2 FedRAMP ISO 27001); Acuvity's own certs not confirmed ↗ 2 | enterprise deployment; runtime inspection of AI/MCP traffic; data handling under Proofpoint's enterprise DLP framework ↗ 2 | on-prem or cloud (Proofpoint enterprise deployment); agentless approach ↗ 2 | Sunnyvale CA (acquired by Proofpoint Sunnyvale) ↗ 2 | Acuvity (acquired by Proofpoint 2025; enterprise AI security) ↗ 2 | internal — undisclosed (enterprise security) ↗ 2 | no public benchmark scores found — vendor does not publish quantitative perf metrics ↗ 2 | not applicable — not a research paper 3 | not applicable — not a research paper 3 | REST (typical SaaS REST + first-party SDK pattern; see vendor docs for language coverage) ↗ 2 | searched not found ↗ 2 | searched not found ↗ 2 | custom ↗ 2 | hard-isolation ↗ 2 | at-rest + in-transit ↗ 2 | SSO + RBAC ↗ 2 | locked ↗ 2 | not specified — vendor docs do not address this dimension ↗ 2 | none ↗ 2 | not applicable 3 | not specified — vendor docs do not address this dimension ↗ 2 | not specified — vendor docs do not address this dimension ↗ 2 | not applicable 3 | not applicable 3 | Yes — Secure MCP servers (Model Context Protocol) is a core capability; protects MCP-based agent integrations ↗ 2 | no data — searched proofpoint.com/us/platform/ai-security, acuvity.ai/, acuvity.ai/integrations/, acuvity.ai/blog/; A2A protocol not advertised ↗ 2 | no data — searched acuvity.ai, proofpoint.com/us/platform/ai-security, docs.acuvity.ai; OpenTelemetry export not advertised ↗ 2 | no data — searched acuvity.ai, proofpoint.com/us/platform/ai-security, docs.acuvity.ai; webhook events not advertised ↗ 2 | Exportable structured audit logs (event lineage, decisions, redactions) for audits and investigations ↗ 2 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | regulated enterprise: financial-services, healthcare, government, legal ↗ 2 | <1day to <1week (enterprise procurement + integration) ↗ 2 | not for hobbyist / non-production use cases ↗ 2 | governance + compliance + audit + adversarial hardening ↗ 2 | BYO agent stack; this layer guards it ↗ 2 | Enterprise DLP for AI products — protects against data exfiltration via memory writes/reads; backed by Proofpoint's enterprise channel. ↗ 2 | Enterprise-tier pricing and complexity; smaller mind-share than Lakera or Microsoft Defender for AI. ↗ 2 | acuvity.ai Proofpoint acquisition ↗ 2 |
| Adaptive-RAG | T3 | Query-complexity routing ↗ ● | vector | similarityhybrid-rerank | long-term | read-only | chunk | inspectable | n/a | Smaller classifier LM predicts query complexity, then routes to no-retrieval / single-step / iterative retrieval as appropriate. NAACL 2024. ↗ ● | Improves overall efficiency and accuracy of QA systems vs adaptive baselines, across query complexity levels. ↗ ● | searched not found ↗ 2 | 2024-04 ↗ ● | no releases ↗ ● | Apache-2.0 ↗ ● | 387★ Jsonnet ↗ ● | searched not found ↗ ● | 449 cites 224/yr ↗ 2 | not applicable — not commercial 3 | not applicable — not commercial 3 | searched not found ↗ 2 | not applicable — not commercial 3 | not applicable — not commercial 3 | searched not found ↗ ● | not applicable — not a company 3 | Seoul National University; Soyeong Jeong Jinheon Baek Sukmin Cho Sung Ju Hwang Jaewoo Kang; NAA… ↗ 2 | searched not found ↗ 2 | no public benchmark scores found ↗ ● | Independently reproduced ↗ 2 | Code public + weights closed ↗ ● | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not documented publicly ↗ ● | none ↗ ● | not applicable 3 | schema with migration ↗ ● | not documented publicly ↗ ● | not applicable 3 | not applicable 3 | not applicable — research paper, no deployed product 3 | not applicable — research paper, no deployed product 3 | not applicable — research paper, no deployed product 3 | not applicable — research paper, no deployed product 3 | not applicable — research paper, no deployed product 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | research, RAG developers (horizontal); some have production deployments ↗ ● | not applicable - research paper or <1hr (OSS reference impl) 3 | most are research papers; production fitness varies by maintainer ↗ ● | retrieval quality / accuracy on long-context QA ↗ ● | BYO LLM + vector store; OSS reference implementations exist ↗ ● | Query-adaptive routing across retrieval strategies — picks the right method per question. ↗ ● | Routing decisions can be wrong; adds inference cost. ↗ ● | GitHub starsuzi/Adaptive-RAG arxiv 2403.14403 ACL Anthology (NAACL) ↗ ● |
| Adept ACT | T3 | Computer-use / workflow-automation agent (Adept now part of Amazon AGI 2024) ↗ 2 | none-trivial | none | session-only | agent-controlled | turn | opaque | out-of-scope | 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)** — co-founders + key team joined Amazon AGI; Adept the company continues with Zach Brock as remaining executive. Important historical entry — ACT models inspired Anthropic Computer Use + OpenAI Operator. ↗ 2 | ACT-1 (2022) and ACT-2 (2023) multimodal action transformers; founded by ex-Google Brain + ex-OpenAI execs; **Amazon AGI acqui-hire June 2024** (David Luan + Maxwell Nye + Kelsey Schroeder joined Amazon AGI) ↗ 2 | not applicable — orchestration platform, not memory product 3 | 2022 (founded); 2022-09 (ACT-1 demo); 2024-06 (Amazon acqui-hire) ↗ 2 | not applicable — orchestration platform, not memory product 3 | not applicable — not OSS 3 | not applicable — no GitHub repo 3 | not applicable — orchestration platform, not memory product 3 | not applicable — not academic 3 | $415M total raised pre-acq (Series B 2023 — General Catalyst, Spark, Greylock); Amazon acqui-hire June 2024 (terms ~$300-500M reported but undisclosed) ↗ 2 | not applicable — orchestration platform, not memory product 3 | Pre-acq: enterprise pilot only ↗ 2 | not applicable — orchestration platform, not memory product 3 | not applicable — orchestration platform, not memory product 3 | not applicable — orchestration platform, not memory product 3 | San Francisco, US ↗ 2 | David Luan (ex-Google Brain), Maxwell Nye (ex-OpenAI), Kelsey Schroeder (CEO post-acqui-hire); ex-Google Brain Niki Parmar early team ↗ 2 | not applicable — orchestration platform, not memory product 3 | not applicable — orchestration platform, not memory product 3 | not applicable — not a research paper 3 | not applicable — not OSS 3 | not applicable — orchestration platform, not memory product 3 | not applicable — orchestration platform, not memory product 3 | not applicable — orchestration platform, not memory product 3 | not applicable — orchestration platform, not memory product 3 | not applicable — orchestration platform, not memory product 3 | not applicable — orchestration platform, not memory product 3 | not applicable — orchestration platform, not memory product 3 | not applicable — not a memory product 3 | not applicable — orchestration platform, not memory product 3 | not applicable — orchestration platform, not memory product 3 | not applicable — not a memory product 3 | not applicable — not a memory product 3 | not applicable — orchestration platform, not memory product 3 | not applicable — not a memory product 3 | not applicable — not a memory product 3 | not applicable — orchestration platform, not memory product 3 | not applicable — orchestration platform, not memory product 3 | not applicable — orchestration platform, not memory product 3 | not applicable — orchestration platform, not memory product 3 | not applicable — orchestration platform, not memory product 3 | not applicable — orchestration platform, not memory product 3 | not applicable — orchestration platform, not memory product 3 | not applicable — orchestration platform, not memory product 3 | not applicable — orchestration platform, not memory product 3 | not applicable — orchestration platform, not memory product 3 | action-transformer (vision + action prediction) ↗ 2 | not applicable — orchestration platform, not memory product 3 | single-vendor (Adept-proprietary ACT models historically) ↗ 2 | not applicable — orchestration platform, not memory product 3 | not applicable — orchestration platform, not memory product 3 | not applicable — orchestration platform, not memory product 3 | not applicable — orchestration platform, not memory product 3 | not applicable — orchestration platform, not memory product 3 | (historical) computer-use action transformers; (post-acq) Amazon AGI roadmap ↗ 2 | Amazon AGI (parent post-acquisition); spiritual predecessor to Anthropic Computer Use + OpenAI Operator + Project Mariner ↗ 2 | Pioneered computer-use action-transformer concept (ACT-1 was first widely-shown demo of LLM controlling desktop apps); strong founding-team pedigree. ↗ 2 | Lost commercial momentum to Amazon acqui-hire June-2024; products effectively shelved (Adept entity continues but no longer shipping new ACT models); inspired but didn't capture the computer-use market. ↗ 2 | www.adept.ai/ ↗ 2 |
| Agency Swarm | T3 | OpenAI-Assistants-based multi-agent ↗ ● | n/a | n/a | n/a | n/a | n/a | n/a | n/a | Agent framework built on OpenAI Assistants API — agencies (roles + comm flow); replaced after Assistants v2 changes. ↗ ● | searched not found ↗ ● | searched not found ↗ ● | 2023 ↗ ● | searched not found ↗ ● | MIT ↗ ● | searched not found ↗ ● | searched not found ↗ ● | searched not found ↗ ● | searched not found ↗ ● | searched not found ↗ ● | searched not found ↗ ● | searched not found ↗ ● | searched not found ↗ ● | searched not found ↗ ● | searched not found ↗ ● | Arsenii (VRSEN) Mordvinov; founded 2024; AI agency / influencer-led OSS framework. ↗ ● | searched not found ↗ ● | searched not found ↗ ● | searched not found ↗ ● | searched not found ↗ ● | searched not found ↗ ● | searched not found ↗ ● | searched not found ↗ ● | not applicable — not a memory product 3 | searched not found ↗ ● | searched not found ↗ ● | searched not found ↗ ● | not applicable — not a memory product 3 | not applicable — not a memory product 3 | not applicable — not a memory product 3 | not applicable — not a memory product 3 | not applicable — not a memory product 3 | not applicable — not a memory product 3 | not applicable — not a memory product 3 | not applicable — not a memory product 3 | searched not found ↗ ● | searched not found ↗ ● | searched not found ↗ ● | searched not found ↗ ● | searched not found ↗ ● | searched not found ↗ ● | searched not found ↗ ● | searched not found ↗ ● | searched not found ↗ ● | searched not found ↗ ● | searched not found ↗ ● | searched not found ↗ ● | searched not found ↗ ● | searched not found ↗ ● | searched not found ↗ ● | searched not found ↗ ● | searched not found ↗ ● | searched not found ↗ ● | searched not found ↗ ● | searched not found ↗ ● | searched not found ↗ ● | searched not found ↗ ● | searched not found ↗ ● |
| Agent KB | T4 | Cross-domain experience KB ↗ 2 | vector | similarity | long-term | | episode | inspectable | llm-arbitrate | Leverages cross-domain experience for agentic problem solving. ↗ 2 | Universal cross-framework agent knowledge base with disagreement gate ensuring retrieved knowledge enhances reasoning. Headline: smolagents +18.7pp at pass@3 (55.2%->73.9%) and OpenHands +4.0pp on SWE-bench (24.3%->28.3%); baseline: same frameworks without Agent KB retrieval; primary datasets: GAIA, Humanity's Last Exam, GPQA, SWE-bench. ↗ 2 | text ↗ 2 | 2025-07 ↗ 2 | not applicable — not OSS 3 | not applicable — not OSS 3 | not applicable — no GitHub repo 3 | 43 cites · 4.3/mo ↗ 2 | 43 cites 43/yr ↗ 2 | not applicable — not commercial 3 | not applicable — not commercial 3 | not applicable — not commercial 3 | not applicable — not commercial 3 | not applicable — not commercial 3 | not applicable — not a deployable product 3 | not applicable — not a company 3 | no data — searched arxiv.org/abs/2507.06229; author affiliations not found in preview ↗ 2 | no scale claim — research ↗ 2 | +18.7pp smolagents pass@3 (GAIA/HLE/GPQA) +4.0pp OpenHands SWE-bench ↗ 2 | No reproductions found — published 2025-07-08 (10 mo ago); no public code ↗ 2 | depth-floor-reached: arXiv 2507.06229 preprint; no official code repo discovered via WebSearch + arXiv abstract scan + GitHub API ↗ 2 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not specified - paper does not address this dimension ↗ 2 | not specified - paper does not address this dimension ↗ 2 | not specified - paper does not address this dimension ↗ 2 | schema with migration ↗ 2 | not specified - paper does not address this dimension ↗ 2 | LLM resolves and overwrites ↗ 2 | not specified - paper does not address this dimension ↗ 2 | not applicable — research paper, no deployed product 3 | not applicable — research paper, no deployed product 3 | not applicable — research paper, no deployed product 3 | not applicable — research paper, no deployed product 3 | not applicable — research paper, no deployed product 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable - research paper 3 | not applicable - research paper 3 | not applicable - research paper 3 | not applicable - research paper 3 | not applicable - research paper 3 | Knowledge-base memory for agent task experience. ↗ 2 | Preprint; KB curation overhead. ↗ 2 | arxiv 2507.06229 ↗ 2 |
| AGENT-RECONFIGURE / Reconfigurable Agent | T4 | Skill-library agent memory ↗ 2 | kv | | long-term | accretion | procedure | plaintext | n/a | Agent that maintains a skill library and reconfigures its toolset per task — procedural memory of successful sub-task solutions. ↗ 2 | searched not found ↗ 2 | searched not found ↗ 2 | 2024-10 ↗ 2 | searched not found ↗ 2 | searched not found ↗ 2 | searched not found ↗ 2 | searched not found ↗ 2 | searched not found ↗ 2 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | searched not found ↗ 2 | not applicable — research paper 3 | searched not found ↗ 2 | searched not found ↗ 2 | searched not found ↗ 2 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | searched not found ↗ 2 | searched not found ↗ 2 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — research paper 3 | not applicable — research paper 3 | searched not found ↗ 2 | searched not found ↗ 2 | searched not found ↗ 2 | searched not found ↗ 2 | searched not found ↗ 2 | searched not found ↗ 2 |
| Agent S | T4 | Open agentic computer-use framework ↗ 2 | hybrid | similarityagentic | long-term | | episode | inspectable | llm-arbitrate | Open agentic framework that uses computers like a human. ↗ 2 | Open agentic framework using computers like a human via experience-augmented hierarchical planning + Agent-Computer Interface (ACI). Headline: 20.58% success rate vs. 11.21% baseline on OSWorld (83.6% relative improvement, +9.37pp absolute); baseline: GPT-4o vanilla OSWorld agent; primary datasets: OSWorld (369 Ubuntu tasks) and WindowsAgentArena. ↗ 2 | text + image (GUI — multimodal desktop/web via screenshots) ↗ 2 | 2024-10 ↗ 2 | v0.3.2 2025-12-16 ↗ 2 | Apache-2.0 ↗ ● | 11.1k★ +545/mo Python ↗ ● | 147 cites · 7.8/mo · 11142* / 1301 forks · 10 influential SS cites ↗ 2 | 146 cites 73/yr ↗ 2 | not applicable — not commercial 3 | not applicable — not commercial 3 | not applicable — research paper 3 | not applicable — not commercial 3 | not applicable — not commercial 3 | not applicable — research paper 3 | not applicable — not a company 3 | Simular Research; no institutional affiliation confirmed ↗ 2 | no scale claim — research (OSWorld benchmark) ↗ 2 | 20.58% OSWorld success +9.37pp OSWorld vs GPT-4o ↗ 2 | Original team only — 1301 forks of https://github.com/simular-ai/Agent-S (mostly fork-only; no public re-implementations identified via WebSearch) ↗ ● | Code public + weights closed — https://github.com/simular-ai/Agent-S · 11142* / 1301 forks · last commit 2026-02 · Apache-2.0 ↗ ● | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not specified - paper does not address this dimension ↗ 2 | not specified - paper does not address this dimension ↗ 2 | not specified - paper does not address this dimension ↗ 2 | not specified - paper does not address this dimension ↗ 2 | not specified - paper does not address this dimension ↗ 2 | LLM resolves and overwrites ↗ 2 | not specified - paper does not address this dimension ↗ 2 | not applicable — research paper, no deployed product 3 | not applicable — research paper, no deployed product 3 | not applicable — research paper, no deployed product 3 | not applicable — research paper, no deployed product 3 | not applicable — research paper, no deployed product 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable - research paper 3 | not applicable - research paper 3 | not applicable - research paper 3 | not applicable - research paper 3 | not applicable - research paper 3 | Computer-use agent with explicit screen-history memory. ↗ 2 | Preprint; screen-history scales aggressively. ↗ 2 | arxiv 2410.08164 ↗ 2 |
| Agent Workflow Memory | T3 | Workflow as memory ↗ 2 | file | injection | cross-session | | skill | inspectable | llm-arbitrate | Workflow-based memory framework component. ↗ 2 | Induces commonly-reused workflows from past experiences and applies them in offline (training) and online (test-time) modes. Headline: +24.6% on Mind2Web and +51.1% relative success-rate on WebArena over baseline; baseline: BrowserGym autonomous agent without human-annotated workflows; primary datasets: Mind2Web and WebArena (812 tasks across 5 websites). ↗ 2 | text (web navigation tasks — Mind2Web + WebArena) ↗ 2 | 2024-09 ↗ 2 | not applicable — not OSS 3 | not applicable — not OSS 3 | not applicable — no GitHub repo 3 | 133 cites · 6.7/mo · 426* / 50 forks · 14 influential SS cites ↗ 2 | 70 cites 35/yr ↗ 2 | not applicable — not commercial 3 | not applicable — not commercial 3 | not applicable — not commercial 3 | not applicable — not commercial 3 | not applicable — not commercial 3 | not applicable — not a deployable product 3 | not applicable — not a company 3 | Georgia Tech + CMU; Zora Zheng et al.; no lead confirmed ↗ 2 | no scale claim — research (1000+ web tasks) ↗ 2 | +24.6% Mind2Web +51.1% WebArena (relative) ↗ 2 | Independently reproduced: 50 forks of https://github.com/zorazrw/agent-workflow-memory; downstream papers cite results ↗ ● | Code public + weights closed — https://github.com/zorazrw/agent-workflow-memory · 426* / 50 forks · last commit 2025-12 · Apache-2.0 ↗ ● | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | strong ↗ 2 | snapshots ↗ 2 | not specified - paper does not address this dimension ↗ 2 | free-form / no schema ↗ 2 | hierarchical ↗ 2 | LLM resolves and overwrites ↗ 2 | not specified - paper does not address this dimension ↗ 2 | not applicable — research paper, no deployed product 3 | not applicable — research paper, no deployed product 3 | not applicable — research paper, no deployed product 3 | not applicable — research paper, no deployed product 3 | not applicable — research paper, no deployed product 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable - research paper 3 | not applicable - research paper 3 | not applicable - research paper 3 | not applicable - research paper 3 | not applicable - research paper 3 | Workflow-pattern abstraction means agents recall multi-step procedures, not just facts. ↗ 2 | Workflow-shaped scope; less applicable to free-form recall. ↗ 2 | OpenReview ↗ 2 |
| Agent2Agent Protocol (A2A) | T2 | Open agent-to-agent protocol ↗ 2 | n/a | n/a | n/a | n/a | n/a | n/a | n/a | Google-led open protocol for agent-to-agent collaboration (April 2025) — agent cards, RPCs, server-sent events. Partner companies: Atlassian, MongoDB, Salesforce, ServiceNow, etc. ↗ 2 | 50+ partner companies at launch. ↗ 2 | searched not found ↗ 2 | 2025 ↗ 2 | searched not found ↗ 2 | Apache-2.0 ↗ 2 | searched not found ↗ 2 | searched not found ↗ 2 | searched not found ↗ 2 | searched not found ↗ 2 | searched not found ↗ 2 | searched not found ↗ 2 | searched not found ↗ 2 | searched not found ↗ 2 | searched not found ↗ 2 | searched not found ↗ 2 | Google + 50+ partner companies; launched 2025-04 as open agent-to-agent protocol. ↗ 2 | searched not found ↗ 2 | searched not found ↗ 2 | searched not found ↗ 2 | Active — last commit 2026-05 · 23,737★ ↗ ● | searched not found ↗ 2 | searched not found ↗ 2 | searched not found ↗ 2 | not applicable — not a memory product 3 | searched not found ↗ 2 | searched not found ↗ 2 | searched not found ↗ 2 | not applicable — not a memory product 3 | not applicable — not a memory product 3 | not applicable — not a memory product 3 | not applicable — not a memory product 3 | not applicable — not a memory product 3 | not applicable — not a memory product 3 | not applicable — not a memory product 3 | not applicable — not a memory product 3 | searched not found ↗ 2 | searched not found ↗ 2 | searched not found ↗ 2 | searched not found ↗ 2 | searched not found ↗ 2 | searched not found ↗ 2 | searched not found ↗ 2 | searched not found ↗ 2 | searched not found ↗ 2 | searched not found ↗ 2 | searched not found ↗ 2 | searched not found ↗ 2 | searched not found ↗ 2 | searched not found ↗ 2 | searched not found ↗ 2 | searched not found ↗ 2 | searched not found ↗ 2 | searched not found ↗ 2 | searched not found ↗ 2 | searched not found ↗ 2 | searched not found ↗ 2 | searched not found ↗ 2 | searched not found ↗ 2 |
| AgentEvolver | T4 | Efficient self-evolving agent system ↗ 2 | vector | similarity | long-term | agent-controlled | skill | inspectable | llm-arbitrate | Towards an efficient self-evolving agent system. ↗ 2 | Self-evolving agent system: self-questioning for task generation, self-navigating for experience reuse, self-attributing for sample efficiency. Headline: superior task-goal completion (avg@8/best@8) over Vanilla GRPO baseline using substantially fewer parameters; baseline: Vanilla GRPO RL without experience guidance/process rewards; primary datasets: AppWorld and BFCL-v3 (multi-turn split). ↗ 2 | text ↗ 2 | 2025-11 ↗ 2 | not applicable — not OSS 3 | not applicable — not OSS 3 | not applicable — no GitHub repo 3 | 30 cites · 5.2/mo · 1427* / 163 forks · Tongyi Lab launch + Threads/X coverage 2025-11 ↗ 2 | 30 cites 30/yr ↗ 2 | not applicable — not commercial 3 | not applicable — not commercial 3 | not applicable — not commercial 3 | not applicable — not commercial 3 | not applicable — not commercial 3 | not applicable — not a deployable product 3 | not applicable — not a company 3 | no data — searched arxiv.org/abs/2511.10395; author affiliations not found in preview ↗ 2 | no scale claim — research ↗ 2 | not applicable — qualitative-only paper (abstract reports "more efficient exploration, better sample utilization, faster adaptation" without specific headline numbers; preliminary experiments) 2 | Too recent — published 2025-11-13 (5.7 mo ago); no reproductions yet expected ↗ 2 | Code public + weights closed — https://github.com/modelscope/AgentEvolver · 1427* / 163 forks · last commit 2026-04 · Apache-2.0 ↗ ● | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not specified - paper does not address this dimension ↗ 2 | not specified - paper does not address this dimension ↗ 2 | not specified - paper does not address this dimension ↗ 2 | schema with migration ↗ 2 | not specified - paper does not address this dimension ↗ 2 | LLM resolves and overwrites ↗ 2 | not specified - paper does not address this dimension ↗ 2 | not applicable — research paper, no deployed product 3 | not applicable — research paper, no deployed product 3 | not applicable — research paper, no deployed product 3 | not applicable — research paper, no deployed product 3 | not applicable — research paper, no deployed product 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable - research paper 3 | not applicable - research paper 3 | not applicable - research paper 3 | not applicable - research paper 3 | not applicable - research paper 3 | Agent that evolves its own memory architecture. ↗ 2 | Preprint; evolutionary stability concerns. ↗ 2 | arxiv 2511.10395 ↗ 2 |
| AgentFold | T4 | Proactive context management ↗ 2 | kv | agentic | session | consolidation | summary | inspectable | llm-arbitrate | Long-horizon web agents with proactive context management. ↗ 2 | Proactive context management via folding inspired by retrospective consolidation; multi-scale granular condensation to deep consolidation. Headline: AgentFold-30B-A3B reaches 36.2% on BrowseComp and 47.3% on BrowseComp-ZH, surpassing DeepSeek-V3.1-671B (30.0% on BrowseComp) and OpenAI o4-mini; baselines: DeepSeek-V3.1-671B, GLM-4.5-355B, OpenAI o4-mini; primary datasets: BrowseComp, BrowseComp-ZH, WideSearch, GAIA. ↗ 2 | text + image ↗ 2 | 2025-10 ↗ 2 | not applicable — not OSS 3 | not applicable — not OSS 3 | not applicable — no GitHub repo 3 | 31 cites · 5.0/mo · 18805* / 1447 forks · Tongyi DeepResearch monorepo (#1 Github trending Oct 2025) ↗ 2 | 30 cites 30/yr ↗ 2 | not applicable — not commercial 3 | not applicable — not commercial 3 | not applicable — not commercial 3 | not applicable — not commercial 3 | not applicable — not commercial 3 | not applicable — not a deployable product 3 | not applicable — not a company 3 | no data — searched arxiv.org/abs/2510.24699; author affiliations not found in preview ↗ 2 | no scale claim — research ↗ 2 | 36.2% BrowseComp 47.3% BrowseComp-ZH ↗ 2 | Original team + downstream attention: 1447 forks of https://github.com/Alibaba-NLP/DeepResearch; published 2025-10-28 (6 mo ago); no full re-implementations identified ↗ ● | Code public + weights closed — https://github.com/Alibaba-NLP/DeepResearch · 18805* / 1447 forks · last commit 2026-02 · Apache-2.0 ↗ ● | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not specified - paper does not address this dimension ↗ 2 | snapshots ↗ 2 | not specified - paper does not address this dimension ↗ 2 | free-form / no schema ↗ 2 | not specified - paper does not address this dimension ↗ 2 | LLM resolves and overwrites ↗ 2 | consolidation/summarization ↗ 2 | not applicable — research paper, no deployed product 3 | not applicable — research paper, no deployed product 3 | not applicable — research paper, no deployed product 3 | not applicable — research paper, no deployed product 3 | not applicable — research paper, no deployed product 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable - research paper 3 | not applicable - research paper 3 | not applicable - research paper 3 | not applicable - research paper 3 | not applicable - research paper 3 | Folded-context agent — compresses old context into summary memory. ↗ 2 | Preprint; folding-quality bounds long-horizon recall. ↗ 2 | arxiv 2510.24699 ↗ 2 |
| Agentic Memory | T4 | Unified short + long-term management ↗ 2 | hybrid | similarityagentic | long-term | consolidation | episode | agent-controlled | llm-arbitrate | Learning unified long-term and short-term memory management. ↗ 2 | Agent-controlled memory operations as tool-based actions for long/short-term memory; trained via three-stage progressive RL with step-wise GRPO. Headline: AgeMem reaches 54.31% avg on Qwen3-4B-Instruct, +23.52% over best baseline (Mem0/A-Mem) and +49.59% over no-memory; baselines: Mem0, A-Mem, LangMem, AgeMem-noRL ablation; primary datasets: ALFWorld, SciWorld, PDDL, BabyAI, HotpotQA. ↗ 2 | text ↗ 2 | 2026-01 ↗ 2 | not applicable — not OSS 3 | not applicable — not OSS 3 | not applicable — no GitHub repo 3 | 15 cites · 3.8/mo ↗ 2 | 14 cites 14/yr ↗ 2 | not applicable — not commercial 3 | not applicable — not commercial 3 | not applicable — not commercial 3 | not applicable — not commercial 3 | not applicable — not commercial 3 | not applicable — not a deployable product 3 | not applicable — not a company 3 | no data — searched arxiv.org/abs/2601.01885; author affiliations not found in preview ↗ 2 | no scale claim — research ↗ 2 | 54.31% ALFWorld/SciWorld/PDDL/BabyAI/HotpotQA avg (Qwen3-4B) +23.52pp vs Mem0/A-Mem ↗ 2 | Too recent — published 2026-01-05 (4.0 mo ago); no reproductions yet expected ↗ 2 | depth-floor-reached: arXiv 2601.01885 preprint; no official code repo discovered via WebSearch + arXiv abstract scan + GitHub API ↗ 2 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not specified - paper does not address this dimension ↗ 2 | snapshots ↗ 2 | not specified - paper does not address this dimension ↗ 2 | not specified - paper does not address this dimension ↗ 2 | not specified - paper does not address this dimension ↗ 2 | LLM resolves and overwrites ↗ 2 | consolidation/summarization ↗ 2 | not applicable — research paper, no deployed product 3 | not applicable — research paper, no deployed product 3 | not applicable — research paper, no deployed product 3 | not applicable — research paper, no deployed product 3 | not applicable — research paper, no deployed product 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable - research paper 3 | not applicable - research paper 3 | not applicable - research paper 3 | not applicable - research paper 3 | not applicable - research paper 3 | Agent-shaped memory with first-class read/write operations exposed to the agent loop. ↗ 2 | Preprint; agent-controlled memory quality bounds value. ↗ 2 | arxiv 2601.01885 ↗ 2 |
| Agentic Plan Caching (APC) | T3 | Test-time plan-template memory ↗ 2 | kv | similarity | cross-session | | skill | inspectable | llm-arbitrate | Extracts, stores, adapts, and reuses structured plan templates from planning stages of agent applications. NeurIPS 2025 poster. ↗ 2 | Plan template extraction from completed agent runs with keyword retrieval and lightweight model adaptation. Headline: 50.31% serving-cost reduction and 27.28% latency reduction on average while preserving 96.61% of optimal performance; baselines: Accuracy-Optimal (large planner LM, no caching) and Semantic Caching; primary datasets: FinanceBench, TabMWP, QASPER, AIME 2024/2025, GAIA. ↗ 2 | text ↗ 2 | 2025-06 ↗ 2 | not applicable — not OSS 3 | not applicable — not OSS 3 | not applicable — no GitHub repo 3 | 11 cites · 1.0/mo ↗ 2 | searched not found ↗ 2 | not applicable — not commercial 3 | not applicable — not commercial 3 | not applicable — not commercial 3 | not applicable — not commercial 3 | not applicable — not commercial 3 | not applicable — not a deployable product 3 | not applicable — not a company 3 | NeurIPS 2025 poster; no author affiliation found at neurips.cc/virtual/2025/poster/11… ↗ 2 | no scale claim — research (NeurIPS 2025) ↗ 2 | 50.31% serving-cost reduction 96.61% of optimal performance ↗ 2 | No reproductions found — published 2025-06-17 (11 mo ago); no public code ↗ 2 | depth-floor-reached: arXiv 2506.14852 preprint; no official code repo discovered via WebSearch + arXiv abstract scan + GitHub API ↗ 2 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not applicable — research paper 3 | not specified - paper does not address this dimension ↗ 2 | not specified - paper does not address this dimension ↗ 2 | not specified - paper does not address this dimension ↗ 2 | free-form / no schema ↗ 2 | not specified - paper does not address this dimension ↗ 2 | LLM resolves and overwrites ↗ 2 | not specified - paper does not address this dimension ↗ 2 | not applicable — research paper, no deployed product 3 | not applicable — research paper, no deployed product 3 | not applicable — research paper, no deployed product 3 | not applicable — research paper, no deployed product 3 | not applicable — research paper, no deployed product 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable — wrong section 3 | not applicable - research paper 3 | not applicable - research paper 3 | not applicable - research paper 3 | not applicable - research paper 3 | not applicable - research paper 3 | Cache-aware planning reduces repeated work for similar tasks. ↗ 2 | Niche use case; cache-invalidation issues common in evolving domains. ↗ 2 | NeurIPS 2025 poster ↗ 2 |