Short version: "memory" for AI agents meant stuffing the chat history back into the prompt until not long ago. In 2026 it is a product category with three distinct designs: extraction-based memory stores that distil conversations into facts (Mem0, Supermemory, Memori), graph memory that keeps entities and relationships with time attached (Graphiti, Cognee, MemOS), and stateful-agent runtimes where memory is part of the agent itself (Letta). Most vendors publish a benchmark where they win. This guide ranks the open-source options with stars, licences and release dates read from the GitHub API on 10 November 2026, explains which design fits which problem, and flags the projects that have stopped moving. Refreshed monthly.
| # | Library | Stars | Licence | Latest release | Design |
|---|---|---|---|---|---|
| 1 | mem0ai/mem0 | 68,738 | Apache-2.0 | ts-v3.3.1 (25 Sep 2026) | Extraction-based memory layer |
| 2 | topoteretes/cognee | 33,201 | Apache-2.0 | v1.6.7 (6 Nov 2026) | Knowledge-graph memory platform |
| 3 | getzep/graphiti | 32,705 | Apache-2.0 | v0.31.2 (20 Oct 2026) | Real-time temporal knowledge graph |
| 4 | supermemoryai/supermemory | 31,583 | MIT | server-v0.0.9 (12 Oct 2026) | Memory and context engine, runs locally |
| 5 | letta-ai/letta | 25,401 | Apache-2.0 | 0.16.8 (14 May 2026) | Stateful agents with self-editing memory |
| 6 | MemoriLabs/Memori | 17,521 | Apache-2.0 (modified) | v3.3.6 (28 May 2026) | LLM-agnostic memory infrastructure |
| 7 | MemTensor/MemOS | 12,343 | Apache-2.0 | v2.0.36 (9 Nov 2026) | "Memory OS" with skill reuse |
| 8 | CaviraOSS/LongMemory | 4,572 | Apache-2.0 | v1.3.0 (20 Dec 2025) | Local store for coding assistants |
1. mem0ai/mem0 (68,738 stars)
Mem0 is the memory layer most people have heard of: a drop-in service that watches an agent's conversations, extracts what is worth keeping and retrieves it next time. Its README leads with a new memory algorithm shipped in April 2026 and a benchmark table claiming 92.5 on LoCoMo and 94.4 on LongMemEval at around 7,000 tokens and roughly one second of latency, a large jump over its own previous numbers. Those are Mem0's own measurements of its managed platform; the README says the open-source SDK lacks some of the optimisations behind them and should expect "directionally similar" results rather than the same numbers. The evaluation harness is public (mem0ai/memory-benchmarks). Apache-2.0, Python and TypeScript SDKs, the TypeScript package at v3.3.1 since 25 September, hosted platform optional. Its old standalone MCP server was archived in March 2026 in favour of a hosted endpoint.
2. topoteretes/cognee (33,201 stars)
Cognee calls itself the free open-source AI memory platform for agents: it turns documents, code and conversations into a self-hosted knowledge graph that persists across sessions, and it is built to work with small models rather than assuming a frontier API. Apache-2.0, v1.6.7 on 6 November, with a community plugin ecosystem. Choose it when memory has to be yours, on your infrastructure, and when the things you remember have structure.
3. getzep/graphiti (32,705 stars)
Graphiti, from the Zep team, builds real-time knowledge graphs for agents with time as a first-class dimension: facts carry when they became true and when they stopped, which is what lets an agent answer "what did the customer want last month versus now". Apache-2.0, v0.31.2 on 20 October. It is the library behind Zep's hosted product, and the one to reach for when memory is really a changing graph of entities.
4. supermemoryai/supermemory (31,583 stars)
Supermemory is a memory and context engine plus an app, MIT licensed and designed to run fully locally. Its README claims first place on LongMemEval, LoCoMo and ConvoMem, and on LongMemEval specifically 95% recall at fifteen results with a 99.4% reduction in context, plus about 50 ms user profiles. As with Mem0, treat vendor benchmarks as a starting point. The server package is young (server-v0.0.9 on 12 October) and the repository is pushed most days (about 150 commits in the last three months).
5. letta-ai/letta (25,401 stars)
Letta, formerly MemGPT, is the original stateful-agent idea: the agent edits its own memory blocks inside its context window and improves over time, and Letta Code now keeps that memory as Markdown files in a git-backed MemFS, with root-level files in the system prompt and subdirectories loaded only when needed. Note the repository change: the README says the current source now lives in letta-ai/letta-code, which contains the agent harness, terminal UI, app server and the runtime behind the desktop and web apps. The letta repository itself last released 0.16.8 in May 2026 and still holds the stars; letta-code (3,811 stars, TypeScript) released v0.34.51 on 10 November. Follow letta-code if you want the active codebase.
6. MemoriLabs/Memori (17,521 stars)
Memori describes itself as agent-native memory infrastructure: an LLM-agnostic layer that turns agent execution and conversation into structured memory. v3.3.6 shipped in May 2026 and the last push was 3 October, so it is slower-moving than the top four. Its LICENSE is the Apache 2.0 text with edited definitions and grant clauses, which is why GitHub does not detect a standard licence, so read it before adopting. The SDK is also cloud-first: the quickstart needs a Memori Cloud API key, with bring-your-own-database and VPC or on-premises deployment as documented options.
7. MemTensor/MemOS (12,343 stars)
MemOS 2.0, code-named Stardust, is a self-evolving "memory operating system": ultra-persistent memory, hybrid retrieval and cross-task skill reuse, built as a Python service with a TypeScript local plugin and a focus on plugging into agent harnesses (its README's newest integration is DeepSeek Harness, with automatic recall, background capture and a local memory viewer). Apache-2.0, v2.0.36 on 9 November; core and plugin releases land every week or two, though v2.0.34 (23 September) to v2.0.36 was a five-week gap. The most ambitious design on this list, and one of the youngest projects on it (created July 2025).
8. CaviraOSS/LongMemory (4,572 stars)
LongMemory is a local-first memory engine whose GitHub description names Claude Desktop, GitHub Copilot, Codex and Antigravity, though its README positions it more broadly, with a Python SDK, HTTP, MCP, a dashboard and a VS Code extension. Apache-2.0. v1.3.0 on GitHub dates from December 2025 (npm has 1.3.3, published 31 August 2026) and the repository was last pushed on 10 November, but activity has thinned: about 40 commits in the last three months, most of them in August. If what you want is "my coding agent remembers my project between sessions" without a platform, this is the self-hosted answer.
The hosted and built-in options
- Zep (getzep/zep, 5,001 stars) has become an enterprise context layer: governed access to company knowledge through what it calls Context Graphs. The open-source part is Graphiti; the repository now holds examples and integrations.
- LangMem (langchain-ai/langmem, 1,759 stars, MIT) provides memory primitives for LangGraph agents: extract facts, refine prompts, maintain long-term memory with any store. No tags or GitHub releases, and only a handful of commits since October, but still maintained (last push 3 November).
- Frameworks with memory built in: agno (43,128 stars) and the memory modules inside LangGraph, CrewAI and the OpenAI Agents SDK cover the simple cases without a separate service. Add one of the libraries above when memory outgrows the framework.
- The MCP memory wave: dozens of small servers (mcp-memory-service and others, each under 3,000 stars) expose a memory tool over MCP, and prompt-and-script hacks such as OptMem (2,387 stars) do the same with no server at all. Useful for a personal setup; too young for a product.
About those benchmarks
Mem0, Supermemory and MemOS each publish LoCoMo or LongMemEval scores, and each is at the top of its own table. The benchmarks are real and public, but the runs are vendor-run, the retrieval budgets differ, and the numbers move with every model release. Use them to shortlist, then evaluate on your own transcripts: the question that matters is whether the agent recalls the right fact from your users' histories at a latency you can afford.
Gone quiet
- kingjulio8238/Memary: last push October 2024.
- memodb-io/memobase: last push and last tag (v0.0.42) January 2026, no GitHub releases.
- mem0ai/mem0-mcp: archived March 2026 (Mem0 now serves MCP from its platform).
How to choose
- Chat product that should remember users: Mem0 or Supermemory; both take minutes to add.
- Memory with structure and time: Graphiti, or Cognee if you also want documents and code in the same graph.
- Agents that manage their own memory: Letta (follow letta-code).
- Coding assistants only: LongMemory, or the built-in memory of your harness.
- Everything must stay in-house: Cognee, Supermemory self-hosted, or MemOS.
FAQ
What is the difference between RAG and agent memory?
RAG retrieves from a corpus you prepared in advance; memory is written by the agent during use and retrieved later, so it changes constantly and is usually per user. Many memory libraries are built on the same vector and graph stores a RAG system uses; the difference is the write path.
Do I need a graph, or is vector memory enough?
Vector memory handles "remember this preference" well. The moment questions involve relationships or change over time ("who approved this, and has that changed?"), a temporal graph such as Graphiti pays for itself.
Which memory library is best for Claude Code or Codex?
For a single developer, LongMemory or Supermemory's local build plug in as local stores; both ship Claude Code and Codex plugins. MemOS's local plugin targets DeepSeek Harness, Hermes and OpenClaw rather than Claude Code or Codex. For a team, Cognee or Mem0 behind an MCP server, so every agent shares one memory.
Are the LoCoMo and LongMemEval scores comparable across vendors?
Only roughly. Each vendor runs its own evaluation with its own retrieval budget and model, so treat the tables as directional and test on your own data before choosing.
How this list is built
Stars, licences, release tags and archive status come from the GitHub API, captured on 10 November 2026; positioning and benchmark claims come from each project's README on the same day and are attributed to the vendor. The top eight are open-source libraries or platforms whose main job is agent memory, actively pushed, with a release in 2026 (LongMemory is included on recent commits despite a 2025 GitHub tag). We re-run the numbers monthly and record any change in this page's changelog. For the graph and vector stores underneath these libraries, see our RAG repositories map.
Comments 0
No comments yet. Be the first to share your thoughts.
Leave a comment