TeleAI-UAGI/Awesome-Agent-Memory — stars, momentum and ranking

Curated systems, benchmarks, and papers etc. on memory for LLMs/MLLMs --- long-term context, retrieval, and reasoning.

🧠 Awesome Agent Memory

A curated map of memory for AI agents — the systems, benchmarks, and research that give LLM and multimodal agents long-term context, persistent recall, and the ability to improve from experience.

👀 Open-source resources (e.g. papers with reproducible code publicly available on Github) are marked in bold font and ranked higher.

🧭 Start Here

If you want to…	Jump to
Add a memory layer to an agent you're building	💿 Products
Choose a benchmark to evaluate a memory system	📏 Benchmarks
Get oriented in the field	📖 Tutorials · 📚 Surveys
Follow research on memory architectures	🔤 Nonparametric Memory · 🔢 Parametric Memory
Build agents that learn from experience	📈 Memory for Agent Evolution
Protect agent memory from poisoning and abuse	🔒 Memory Security & Defense

📋 How this list is curated

Open-source products are ordered by GitHub star count — an objective, CI-checked popularity signal, not a quality ranking or an endorsement. Products with fewer than 100 stars sit in a collapsed Emerging projects section and graduate into the main list once they cross that threshold.
Bold marks resources with reproducible code publicly available.
Descriptions are factual, not promotional (see the contributing guide).
This list is maintained by Bloo-Mind AI and the Ubiquitous AGI team at TeleAI. Entries affiliated with the maintainers are marked with † and follow the same ranking, format, and style rules as every other entry. One convention applies to all entries regardless of affiliation: an API-compatible drop-in replacement for a listed product is nested under that product as an unranked sub-item rather than given its own star-ranked position.
Projects that are inactive, archived, or whose claims are disputed move to the Archival subsection with a neutral status label, links to the evidence, and the date the status was last checked.

📰 In the News

🗂️ Table of Contents

[💿 Products](#-products)
[📖 Tutorials](#-tutorials)
[📚 Surveys](#-surveys)
[📏 Benchmarks](#-benchmarks)

    [💬 Plain-Text Benchmarks](#-plain-text-benchmarks)
    [🎬 Multimodal Benchmarks](#-multimodal-benchmarks)
    [🎮 Dynamic Benchmarks & Simulation Environments](#-dynamic-benchmarks--simulation-environments)

[🔤 Papers - Nonparametric Memory](#-papers---nonparametric-memory)

    [📝 Text Memory](#-text-memory)
    [🌐 Graph Memory](#-graph-memory)
    [🎥 Multimodal Memory (for Understanding)](#-multimodal-memory-for-understanding)
    [🎥 Multimodal Memory (for Generation)](#-multimodal-memory-for-generation)

[🔢 Papers - Parametric Memory](#-papers---parametric-memory)
[📈 Papers - Memory for Agent Evolution](#-papers---memory-for-agent-evolution)

    [🧭 Reinforcement Learning & Continual Learning](#-reinforcement-learning--continual-learning)
    [🧩 Context Engineering & Harness Engineering](#-context-engineering--harness-engineering)

[🔬 Papers - Memory in Cognitive Science](#-papers---memory-in-cognitive-science)
[🔒 Memory Security & Defense](#-memory-security--defense)
[📰 Articles](#-articles)
[👥 Workshops](#-workshops)

If you find this page helpful, please give it a ⭐️ — starring also keeps updates in your GitHub feed.

🤝 Contributions welcome! Feel free to open an issue or submit a pull request to add papers, fix links, or improve categorization — see the contributing guide for entry formats.

💿 Products

Open-Source

Ordered by the number of GitHub stars. Products with fewer than 100 stars continue the list inside the collapsed Emerging projects section below — they graduate into the main list once they cross that threshold. An API-compatible drop-in replacement for a listed product appears as an unranked sub-item under that product.

Claude-Mem [code] [docs] [blog] Hook-based session capture compressed into searchable observations and re-injected into later sessions across Claude Code, Codex, Cursor, OpenClaw and other hosts; optional hosted cloud sync.
Mem0 [code] [docs] [paper] [blog] Universal memory layer for AI agents.
- TeleMem † [code] [docs] [paper] API-compatible high-performance drop-in replacement for Mem0 (import telemem as mem0); listed as an unranked sub-item of Mem0 per the drop-in replacement convention. Maintainer-affiliated.
Hindsight [code] [paper] Agent memory layer that learns from interaction feedback to improve recall over time.
OpenViking [code] [docs] [paper1] [paper2] [blog] Self-evolving context database for AI agents that unifies agent memory, knowledge RAG, and skills behind one storage/retrieval layer, with an MCP server for cross-session read/write.
Zep (powered by Graphiti) [code] [paper] [blog] Real-time temporal knowledge graphs for AI agents.
Cognee [code] [paper] [blog] Memory engine that ingests data into a hybrid graph + vector knowledge graph for cross-session agent recall.
gbrain [code] Garry's opinionated OpenClaw/Hermes agent brain.
agentmemory [code] Persistent memory for AI coding agents.
TencentDB Agent Memory [code] Fully local long-term memory for AI agents via a 4-tier progressive pipeline, with zero external API dependencies.
Letta (formerly MemGPT) [code] [paper] [research] [blog] Stateful-agent platform with hierarchical memory that learns and self-improves over time.
Second Me [code] [paper] Personal AI trained on the user to represent them across applications.
MemU [code] [blog] Memory layer for 24/7 proactive agents.
EverOS (part of EverMind) [code] [blog] Toolkit for building, evaluating, and integrating long-term memory in self-evolving agents.
MemOS (by MemTensor) [code] [paper] Memory OS for LLM agents with hybrid retrieval and cross-task skill reuse.
Honcho [code] [research] [blog] [eval] Memory library for stateful agents with a focus on user modeling.
MemoryBear [code] [paper] Memory framework providing human-like episodic and semantic recall to AI agents.
engram (by Gentleman-Programming) [code] Persistent memory for AI coding agents — agent-agnostic single Go binary with SQLite + FTS5, exposed via MCP server, HTTP API, CLI, and TUI.
OpenMemory [code] Local persistent memory store for LLM apps (Claude Desktop, Copilot, Codex, etc.).
memory-lancedb-pro [code] [blog] [video] Enhanced LanceDB memory plugin for OpenClaw
MIRIX [code] [paper] [blog] Multi-agent personal assistant that captures on-screen activity and consolidates it into structured memory.
MemMachine [code] [blog] Interoperable memory layer providing extensible storage and retrieval primitives for AI agents.
Memobase [code] User profile-based long-term memory for AI chatbot applications.
Memanto [code] [paper] [docs] Typed semantic memory with remember/recall/answer operations and information-theoretic retrieval.
LangMem [code] [blog] LangChain's memory primitives for storing, recalling, and managing agent state in LangGraph workflows.
Omnigraph [code] Object-storage-native graph engine for agent memory with git-style branch/merge workflows.
PowerMem [code] Persistent, self-evolving memory for AI agents — hybrid vector/full-text/graph retrieval with LLM-driven extraction, Ebbinghaus-style decay, and two-layer Experience + Skill distillation; from the OceanBase team.
Mem9 [code] [blog] Local private memory hub for OpenClaw and similar coding agents.
Puppyone [code] [docs] Filesystem-shaped agent memory with auto-versioning, per-agent ACLs, and data connectors; accessible via MCP/REST/CLI.
deja [code] Indexes the session transcripts twenty coding agents already write to disk, retroactively — local BM25 recall over them, with credentials redacted at index time.
CodeAlmanac [code] Repo-local Markdown wiki for AI coding agents that preserves project conversations, decisions, and implementation context.
projectmem [code] [docs] [paper] Local-first, event-sourced memory for AI coding agents: an append-only event log served via MCP, plus a pre-commit gate that warns before repeating a failed fix.
Memorix [code] Local-first cross-agent memory layer for coding agents via MCP — SQLite-backed project memory with observation, reasoning, and git-derived fact types, plus task-lensed context briefs.
HMS (Holographic Memory System) [code] Long-term memory QA framework that wraps OpenAI clients with automatic recall and retain, PostgreSQL-backed, evaluated on LongMemEval.
Vestige [code] [release] Local-first cognitive memory MCP server for coding agents, with FSRS-6 decay, spreading activation, active suppression, Receipt Lock, and an inspectable dashboard.
Compartment [code] Offline, encrypted-at-rest vector memory for agents via MCP server, Python, or CLI; AEAD-encrypted embeddings, hybrid recall, per-record crypto-shred deletion, hash-chained audit log.
Caura (formerly MemClaw) [code] [blog] Governed shared memory for AI agent fleets — cross-agent knowledge sharing with permissions, audit trails, and self-learning.
MisakaNet [code] [wiki] Git-based distributed swarm memory; agents share lessons across nodes via GitHub Issues.
Statewave [code] [docs] [blog] Open-source memory runtime for AI agents serving reproducible, provenance-tagged context bundles instead of query-time retrieval; self-hosted on Postgres + pgvector with Python/TypeScript SDKs.
Mnemory [code] Multi-type agent memory (facts, preferences, episodic) with TTLs, user/agent scoping, and an MCP server.
OMEGA [code] [blog] MCP server exposing 25 memory tools for AI coding agents.
Belief Context Graph [code] [docs] Confidence-aware belief graph organizing long-horizon agent context into trackable, updatable belief states; includes SDK, graph-building tools, visualization UI, and agent benchmark.
Remnic [code] [docs] [paper] Local-first Markdown memory shared across coding agents and MCP clients, with per-result provenance, correction workflows, and the MemCorrect benchmark.
Memov [code] Git-based, traceable memory layer for Claude Code.
CommonGround Kernel [code] PostgreSQL-backed shared work-record substrate for human-agent and multi-agent systems, with durable handoff facts, causal lineage, and pull-first recovery across runtimes.

🌱 Emerging projects — open-source products with fewer than 100 GitHub stars, same format and ordering (click to expand)

causal-memory [code] [eval] Local-first agent memory in Rust: facts and typed decision→outcome causal edges (caused/enabled/prevented) on one SQLite store, with inhibitory spreading activation, MCP server, CLI, Python bindings.
taOSmd [code] [eval] Local-first, offline agent memory: an append-only transcript yields a typed temporal knowledge graph with source-grounded, verifier-checked facts and hybrid retrieval, tuned for small local models.
Wenlan [code] [docs] Local-first AI knowledge base and LLM wiki that distills agent work into source-cited pages and serves them to MCP clients.
Synap [code] [docs] Long-term memory layer that extracts facts, preferences, episodes, and temporal events from conversations; integrates with most major agent frameworks.
LWC [code] Agent-operated persistent memory CLI with source-cited Wiki pages, full-text search, document graphs, and CodeGraph indexes for cross-session project recall.
RetainDB [code] Local-first agent memory with noise filtering, typed facts, reuse-based reinforcement, and BM25 + vector + graph retrieval with RRF and reranking; Apache-2.0 core, BSL-1.1 server.
widemem-ai [code] Lightweight memory layer with importance scoring, temporal decay, and 3-tier hierarchy.
memclaw (Felo) [code] Persistent project memory for AI coding agents — isolated per-project workspaces, a web dashboard to review what the agent remembers, and team collaboration.
Mi-Memory [code] [docs] [paper] Lifecycle memory framework for personal assistant agents from Xiaomi's Darwin Agent team; multi-source user state with provenance, editable correction and forgetting, device-adaptive deployment.
Data Olympus [code] Governed project memory for AI coding agents: agents propose learnings, humans promote them, and MCP retrieval serves only in-force knowledge after validity and supersession checks.
Mnemoverse [code] Persistent memory API for agents over MCP: scores importance on write, builds Hebbian associations, and re-ranks recall from outcome feedback; managed engine, open MCP client.
ZenBrain [code] [paper] Neuroscience-inspired 7-layer memory architecture for autonomous agents in zero-dependency TypeScript, with FSRS spaced repetition, Hebbian learning, and sleep-cycle consolidation.
InvMem [code] Vanilla RAG baseline (chunking, embeddings, FAISS/SQLite vector store) behind a synchronous Add/Search API; #1, Agent Memory Leaderboard (academic textual, 2026-08).
Selvedge [code] [docs] Local decision memory for coding agents: records code decisions, rejected approaches and revisit conditions in SQLite; retrieved through MCP or CLI.
Fidelis Memory [code] [eval] Local-first memory for coding agents: MCP recall returns original passages verbatim via vector search, with an explicit BM25/RRF hybrid mode; no generative LLM in retrieval.
GoodMemory [code] [docs] Local-first, auditable memory layer for AI agents and coding hosts, with durable SQLite, embedding-free recall, MCP access, and opt-in governed writeback.
Tree Ring Memory [code] Local-first memory lifecycle for AI agents with a Rust CLI, SQLite/FTS recall, audit, forgetting, consolidation, and Ratatui TUI.
A3M Router [code] Multi-model LLM router with persistent memory (MemoryTree), cross-session context-window management, conversation memory with semantic recall, and ObsidianVault integration.
ReFind [code] Memory retriever that plans iterative searches over a conversation-level BM25 index and returns contextual evidence blocks; #2, Agent Memory Leaderboard (academic textual, 2026-08).
Lians agent memory [code] [eval] Bitemporal agent memory with deterministic supersession, point-in-time recall, MCP access, audit trails, and local SQLite or PostgreSQL storage.
Agentic Task System [code] Agent-native context layer over your existing task app (TickTick; Notion/Obsidian planned), exposing hybrid retrieval over tasks/notes to agents via a CLI with pluggable storage adapters.
myc [code] [docs] Local task-and-memory layer for coding agents: decision oplog with hybrid search, session/repo-scoped recall, PreCompact hook saving the episode before compaction; Bun + SQLite.
FluctlightDB [code] [paper] Embedded database engine for AI agents with experience()/activate() API and reproducible LoCoMo evaluation.
AtMem [code] Local-first agent memory with governed retrieval, provenance, lifecycle controls, delegated context delivery, execution evidence, SQLite storage, MCP, and an audit dashboard.
Lint-AI [code] Agent memory and semantic review engine with lexical, temporal, and graph-aware retrieval across conversations, documents, code, and agent sessions.
Lockstep [code] Shared decision ledger for teams using AI coding agents: records accepted decisions, rationale, and rejected options, and briefs new agent sessions via MCP.
chamnan [code] [data] Repository-local context for coding agents: an architecture index, impact map, and decision records committed beside the code.
elephant [code] Claude Code plugin keeping per-repo memory in a committed ELEPHANT.md plus a global cross-repo file; hooks load both at session start and prompt saves.
hermeneutic [code] Turns corrections from AI agent logs into guidance for similar tasks: local semantic memory with prompt-context hooks and response checks.
sqlite-graph-memory [code] Graph RAG memory for agents over a markdown vault: dense retrieval, hand-curated wikilink 1-hop expansion, cross-encoder rerank, per-turn SQLite ledger.
ActiveMemoryIndex [code] Dual store of verbatim timestamped turns and atomic first-person facts, retrieved in the same register; #3, Agent Memory Leaderboard (academic textual, 2026-08).
inspeximus (formerly mnemo) [code] Zero-dependency memory layer and MCP server with value-ranked recall, per-type decay, keyed supersession, revert-based correction, signed provenance, tamper-evident receipts, and cross-store erasure.
Graphmem [code] Local-first memory for coding agents in Rust: scoped memories and an entity graph in SQLite, recalled via local embeddings plus Personalized PageRank; MCP server.
Synapse [code] Zero-dependency Markdown memory vault with ChatGPT/Claude imports, SQLite keyword and optional semantic search, linked wiki pages, MCP retrieval, and human-approved agent notes.
skillmem [code] Skill memory for Claude Code and Codex agents: records how tasks were solved, recalls them, reinforces what works, decays the rest; SQLite + FTS5 MCP server.
Hyperconsciousness [code] Encrypted, append-only knowledge store for humans and agents: signed records sync across devices and are exposed through scoped, expiring grants over MCP.
archon-memory-core [code] Local-first agent memory with nightly consolidation, active forgetting, and salience scoring.
birkin-mnemosyne [code] Stdlib-only Python memory for agents: Markdown vault with BM25, Korean bigrams and usage decay; model curation passes through a deterministic safety-clamping executor; optional MCP server.
Agent Knowledge Cycle [code] [paper] Six-phase knowledge cycle specification (ADRs, JSON schemas, reference implementation) that turns coding-agent sessions into persistent skills, rules, and memory.
kgai [code] Local-first immutable knowledge graph of engineering decisions for AI coding agents; superseded decisions and rejected approaches stay queryable; embedded graph DB, opt-in team sync.
Talamus [code] [docs] [eval] Local-first agent memory that stores source-grounded Markdown, preserves bitemporal history and provenance, and exposes search, recall, and review-gated correction through MCP.
kannaka-memory [code] [eval] Rust agent memory as a wave-interference medium: phase-coded wavefronts, bilateral hemispheres, dream consolidation and forgetting, NATS swarm sync; CLI and MCP plugin.
agent-memory-doctor [code] Boot-time integrity checker for agent memory files and SQLite session stores: fingerprint and hash checks, size and freshness limits, archive verification; CLI and GitHub Action.
Mnemosyne [code] Memory engine for OpenClaw and Hermes: JSONL and Markdown layers ranked by weighted importance, recency, keyword and hit-frequency scoring; offline by default, DashScope embeddings opt-in.
PackRat [code] Auto-learning codebook compression that shrinks agent context files while keeping them LLM-readable.
Hybrid Episodic Memory [code] Deterministic weighted reciprocal-rank fusion of BM25 and dense retrieval over raw conversational turns, no LLM in the path; #6, Agent Memory Leaderboard (academic textual, 2026-08).
Verified Memory Vault [code] Obsidian vault doubling as Claude Code memory: deterministic health-score linter (undated entries, duplicates, dead links) plus a git pre-commit hook refusing mass deletions.
memgres [code] Versioned document memory for AI agents over one Postgres; lexical or semantic recall, diff-based history, git-blame line attribution, GDPR-erasable, multi-tenant via MCP/HTTP.
FlowGrid AML Retriever [code] Deterministic, evidence-first Add/Search retriever that stores every original message and returns ranked, traceable source evidence; #8, Agent Memory Leaderboard (academic textual, 2026-08).
engram (by FBISiri) [code] Go memory service on Qdrant with write-time dedup and importance gating, type-based TTL decay, reflection into insights, and MCP plus REST interfaces.
AML Memory MVP [code] Evidence-only, typo-tolerant retriever over English and CJK text using SQLite FTS5, character n-grams, and conversation-neighbor expansion; #10, Agent Memory Leaderboard (academic textual, 2026-08).
ExperienceNet [code] [docs] Self-hosted experience network for agents: search and write real technical attempts (problem/conditions/action/outcome), lexical + semantic hybrid retrieval over PostgreSQL/pgvector, gap capture, reuse feedback.
claude-memory-tidy [code] Maintenance layer for always-loaded agent memory files: deterministic budget guard, orphan-note coverage, and verbatim folding into warm sub-indexes, guarding against silent truncation.
YYLO Ledger [code] Git-native task and Record store for coding-agent workflows: reviewable current state, append-only history, dependency-aware work, and bounded queries.
Akephalos [code] Local-first, markdown-based portable agent profile (preferences, rules, durable memories) synced across agents via plain files and Git.
溯忆 (Suyi) [code] Dual-temporal memory engine for AI agents — SQLite-backed, zero-dependency, Ebbinghaus-decayed fact storage with skill crystallization.
Panella [code] Self-hosted governed memory over MCP; agent writes become durable only after human approval with verifiable receipts; Apache-2.0.
Hybrid Memory Search [code] Local hybrid-retrieval memory service with per-user isolation, synchronous embedding on write, and fused BM25/dense ranking; #4, Agent Memory Leaderboard (academic textual, 2026-08).
ChronoHybridMem [code] Evidence-only textual memory over SQLite FTS5 with optional LLM fact extraction and multi-route candidate recall; #5, Agent Memory Leaderboard (academic textual, 2026-08).
Chronicle Memory [code] Evidence-only memory over SQLite with a hybrid lexical scorer adding phrase, temporal, and session-diversity bonuses; #7, Agent Memory Leaderboard (academic textual, 2026-08).
MemoryAgent [code] FastAPI + ChromaDB memory system with local sentence-transformers embeddings and similarity-based write deduplication; #9 as LLLMemoryAgent, Agent Memory Leaderboard (academic textual, 2026-08).
FeedRecall [code] Local-first MCP memory for saved social discoveries, with source dates, project relevance, evidence lifecycle, timelines, and bounded recall for coding agents.
RCLL [code] [docs] [eval] Self-hosted shared memory for a fleet of agents: topic rooms, L0–L3 depth, Postgres/pgvector; the read path invokes no language model. Fork of Hindsight.
notebook.py [code] Single-file Python tool for Claude Code transcript memory, stored as editable Markdown with source references and retained history of retired entries.
kith [code] [eval] Relationship memory between agents on SQLite: append-only observations derive trust, reliability, and sentiment views with per-entry visibility scopes.
alethech [code] Verifiable agent continuity in Python: Ed25519-signed memory commits in a hash-linked DAG, identity-preserving key rotation, encrypted portable history; local-first, zero-LLM.

Closed-Source

MemoraX [spec] [platform] Memory layer for long-horizon agents from MemoraX AI; core system is API-only; #1, Agent Memory Leaderboard (industry textual, 2026-08).
- MemoraX Code
  [code] Coding-agent memory product—not a text-chatbot memory layer—that carries engineering experience, repository knowledge, preferences, and procedures across tasks and sessions.
MemoryLake [blog]
Supermemory [code] [docs] Memory API, hosted or free self-hosted; SDKs, MCP server, and dashboard are MIT, but the memory engine ships only as a prebuilt binary.
Memories.ai [research] [paper] [blog]
Macaron Mind Lab [blog] [paper]
Mem 2.0 [blog]
M-Flow
TwinMind [blog]
Penfield [blog]
Sonzai
Threadline [partial-code] [schema] [docs]
Remio Local-first personal knowledge base that indexes files, webpages, recordings, notes, emails, and messages for agent retrieval via search and RAG.
AccInt [partial-code] Local-first MCP Work Model for coding agents that retrieves scored memory, records actions, and credits real outcomes; engine is a closed-source binary.
Agentage Memory Remote MCP memory server (OAuth 2.1 + PKCE + DCR) giving Claude, Cursor, and ChatGPT one shared markdown memory mirrored locally as files you own.
screenpipe [source-available] [license] [docs] Local-first work memory that captures screen, audio, input, browser, and meeting context for search and agent retrieval.
Firekeep [source-available] [license] [docs] Self-hosted shared memory, working context, coordination, and evidence for Claude Code, Codex, Kiro, OpenCode, and other MCP clients.
ORANO [docs] [blog] Consumer app that distills saved Reels, videos, articles, and PDFs into projects and memory facts that the user's agent reads over a read-only MCP server.
Wontopos (Tablet 2) [paper] [eval] Memory API with no language model in the retrieval path; paper-reported 95.7% LongMemEval-S, 95.2% recall@5 over 70 language pairs, and BEAM-1M 67.5% with published harness.
Perseus Vault (formerly Mimir) [docs] Local MCP memory server as a single Rust binary: SQLite FTS5 plus vector hybrid search, AES-256-GCM at rest; GitHub source repository not publicly accessible (2026-09).
ContextStream [partial-code] [docs] [eval] Hosted MCP context layer for coding agents: persistent decisions and lessons, semantic code search, post-compaction recovery; MIT-licensed Rust client, hosted backend closed-source.
Sentra Organization-wide memory layer for teams and agents: ingests meetings, mail, tickets, and repositories into a bi-temporal fact graph served over REST and MCP.
Moraine Home [source-available] [license] Local-first memory workbench for personal agents and companions: hybrid keyword/vector retrieval, human-and-agent review of candidates, reversible consolidation, timelines, and MCP/HTTP interfaces.
ByteRover [source-available] [license] [paper] [docs] LLM-curated hierarchical context tree for coding agents, with git-like branching, cloud sync, and MCP; formerly Cipher.
Context Mode [source-available] [license] Context-window optimization for AI coding agents: diverts large tool outputs into a locally searchable store and persists session memory across platforms via MCP and hooks.
Agent QA [source-available] [license] QA agent that retains persistent test memory to reuse prior runs and self-heal natural-language web and mobile tests.
RE-call [source-available] [license] [docs] [eval] Postgres plus pgvector memory retrieval for AI agents, with provenance, trust verdicts, tenant isolation, MCP access, and abstention when evidence is insufficient.
past.dev [docs] [eval] Temporal memory API: stores timestamped text with its source; recall returns facts current at a given time and what each replaced; REST and remote MCP.

Archival

Projects that are inactive or whose claims have been disputed by third parties. Status labels link to the evidence and note when the status was last checked.

MemPalace ⚠️ Disputed (third-party critiques challenge the project's core claims; last checked 2026-07) [code] [critique1, critique2] Developed by actress Milla Jovovich and her friends
Memvid ⚠️ Disputed (technical critique raised in GitHub issues, since deleted but archived; last checked 2026-07) [code] [critique (archived)]
Memary ❄️ Inactive (no significant development activity; last checked 2026-07) [code]

📖 Tutorials

🗓️ 2026

Agent Memory Techniques (NirDiamant): 30 runnable Jupyter notebooks covering conversation buffers, vector stores, knowledge graphs, episodic and semantic memory, Mem0, MemGPT/Letta, Zep, Graphiti, and LoCoMo benchmarks [code]
Choose an agent-memory architecture (sir-ad): Decision guide mapping four memory jobs to five architecture patterns, minimum controls, evaluation baselines, and primary research.
Tools, Actions, Memory, and Context †: Chapter 4 of the textbook Multi-Agent Systems: A Contemporary Treatment.

🗓️ 2025

ACM SIGIR-AP 2025 Tutorial: Conversational Agents: From RAG to LTM † [paper] [code]
Daily Dose of DS: A Practical Deep Dive Into Memory Optimization for Agentic Systems [Part-A] [Part-B] [Part-C]

📚 Surveys

🗓️ 2026

🗓️ 2025

🗓️ 2024

A Survey on the Memory Mechanism of Large Language Model based Agents [code]

📏 Benchmarks

💬 Plain-Text Benchmarks

🗓️ 2026

Agent Memory Leaderboard [code] Public evaluation platform: participants expose Add/Search APIs and are scored on textual-memory and coding-agent-memory tracks.
Agent Memory Benchmark (AMB) [code] Open harness and leaderboard scoring memory providers on accuracy, latency, and token cost over six datasets; built by Vectorize, whose Hindsight is among the providers.
OmniMemEval [code]
Are We Ready For An Agent-Native Memory System? (The MemoryData Paper) [code]
Locomo-Plus: Beyond-Factual Cognitive Memory Evaluation Framework for LLM Agents [code]
LoCoMo Refined: Recalibrating LoCoMo with Stricter LLM Judging and A Cleaned Dataset [code]
Agent-Memory Integrity Benchmark [code]
Cost and Accuracy of Long-Term Memory in Distributed Multi-Agent Systems Based on Large Language Models [code]

🎬 Multimodal Benchmarks

🗓️ 2026

🗓️ 2025

🗓️ 2024

🗓️ 2023

🎮 Dynamic Benchmarks & Simulation Environments

🗓️ 2026

EmbodiedMemory-Bench: Benchmarking Embodied Memory for Long-Horizon Embodied Tasks [code] [data]
SEAGym: An Evaluation Environment for Self-Evolving LLM Agents [code]
Beyond Static Dialogues: Benchmarking Realistic, Heterogeneous, and Evolving Long-Term Memory [code] [data] [proj]
AMemGym: Interactive Memory Benchmarking for Assistants in Long-Horizon Conversations [code] [proj]
StreamMemBench: Streaming Evaluation of Agent Memory for Future-Oriented Assistance
agent-memory-bench [code] [data] Preregistered harness scoring memory layers for coding agents by executing task checkers rather than judging text; the author's own RE-call is among the arms.

🗓️ 2025

MemoryBench: A Benchmark for Memory and Continual Learning in LLM Systems [code] [data]
ARE: Scaling Up Agent Environments and Evaluations (The Gaia2 Paper) [code]

🗓️ 2024

AppWorld: A Controllable World of Apps and People for Benchmarking Interactive Coding Agents [code]

🔤 Papers - Nonparametric Memory

📝 Text Memory

🗓️ 2026

🗓️ 2025

🗓️ 2024

🗓️ 2023

RET-LLM: Towards a General Read-Write Memory for Large Language Models

🌐 Graph Memory

🗓️ 2026

🗓️ 2025

🗓️ 2024

🎥 Multimodal Memory (for Understanding)

🗓️ 2026

🗓️ 2025

🗓️ 2024

🎥 Multimodal Memory (for Generation)

🗓️ 2026

🗓️ 2025

🔢 Papers - Parametric Memory

🗓️ 2026

🗓️ 2025

🗓️ 2024

🗓️ 2023

📈 Papers - Memory for Agent Evolution

🧭 Reinforcement Learning & Continual Learning

🗓️ 2026

🗓️ 2025

🧩 Context Engineering & Harness Engineering

🗓️ 2026

🗓️ 2025

🔬 Papers - Memory in Cognitive Science

🗓️ 2026

🗓️ 2025

🗓️ 2024

A Generative Model of Memory Construction and Consolidation

🔒 Memory Security & Defense

Agent Memory Guard [code] OWASP runtime defense layer that screens agent memory writes for poisoning: multi-layer validation with semantic anomaly detection, entropy scoring, and provenance verification.
From Untrusted Input to Trusted Memory: A Systematic Study of Memory Poisoning Attacks in LLM Agents (The MPBench Paper)
inspeximus (formerly mnemo) poisoning probes [attack] [defense] Runnable probe scripts demonstrating that provenance written into a memory record is forgeable, and that a retrieval-time corroboration gate raises the cost of memory-poisoning attacks.

📰 Articles

🗓️ 2026

Harness Engineering for Self-Improvement
221 Agents: Multi-Agent Coordination Lessons (archived; original site currently unreachable)
OpenClaw Multi-Tenancy: Why VM-Per-User Does Not Scale (archived; original site currently unreachable)

🗓️ 2025

Survey of AI Agent Memory Frameworks

🗓️ 2024

🗓️ 2023

LLM Powered Autonomous Agents

👥 Workshops

🗓️ 2026

🗓️ 2025

ACL 2025: The First Workshop on Large Language Model Memorization (L2M2) [proceedings]

📑 Citation

To cite this collection itself, use the metadata in CITATION.cff (GitHub's "Cite this repository" button), or:

@misc{zhang2025awesomeagentmemory,
        author = {Zhang, Dell and Sun, Changzhi and Luo, Jixiang and Chen, Xiangyu and Li, Xuelong},
        title = {Awesome Agent Memory: Curated Systems, Benchmarks, and Papers on Memory for {LLMs}/{MLLMs}},
        year = {2025},
        howpublished = {\url{https://github.com/TeleAI-UAGI/Awesome-Agent-Memory}}
}

This list grew out of the maintainers' SIGIR-AP 2025 tutorial, which you can cite as the related publication:

@inproceedings{zhangConversationalAgentsRAG2025,
        author = {Zhang, Dell and Feng, Yue and Liu, Haiming and Sun, Changzhi and Luo, Jixiang and Chen, Xiangyu and Li, Xuelong},
        title = {Conversational Agents: From {RAG} to {LTM}},
        year = {2025},
        isbn = {9798400722189},
        doi = {10.1145/3767695.3769671},
        booktitle = {Proceedings of the 2025 Annual International ACM SIGIR Conference on Research and Development in Information Retrieval in the Asia Pacific Region (SIGIR-AP)},
        pages = {447–452},
        location = {China}
}

Star History

Regenerated weekly by a scheduled workflow; the previous star-history.com live chart broke when GitHub restricted the stargazers API to repo admins and collaborators in June 2026.

If you find this page helpful, please give it a ⭐️ — starring also keeps updates in your GitHub feed.

Made with ❤️ by Bloo-Mind AI Ltd and the Ubiquitous AGI team at TeleAI.

[![Bloo-Mind Logo](assets/bloo-mind.png)](https://www.bloo-mind.ai/)
&nbsp;&nbsp;&nbsp;
 ![TeleAI Logo](assets/TeleAI.png)