Best GitHub Repos for AI Agent Memory (2026): 10 Open-Source Projects Compared

Quick answer: The best GitHub repos for AI agent memory in 2026 are mem0, MemPalace, Hindsight, Cognee and Graphiti.

All ten here are open source. Pick by use case, and check the licence before you build on one.

AI agent memory is the layer that lets an agent remember facts, preferences and past work between sessions. Without it, every conversation starts from zero.

This list covers the most-starred open-source memory projects, plus the vector databases underneath them.

The 10 best AI agent memory repos at a glance

Star counts come from the GitHub API on November 2, 2026. They show popularity, not quality, and the order reflects fit and maturity rather than stars alone.

How we picked these repos

We searched GitHub for agent-memory projects and kept the ones that are actively maintained and built for memory. Then we weighed four things.

• Fit for the job: it is built for agent memory, not only general search.
• Upkeep: recent commits, and not archived.
• Licence: a clear open-source licence you can read.
• Evidence: documentation, a paper or published results.

We have not run these projects ourselves. Details come from each repository's README and GitHub data.

The 10 best GitHub repos for AI agent memory

1. mem0

Best for: adding a memory layer to almost any agent or app.

mem0 describes itself as the memory layer for AI agents. It offers an open-source library and a managed platform.

Why it ranks first: it is the most-starred dedicated memory project and was last updated on October 30, 2026.

Heads-up: its README reports benchmark scores run on its managed platform, so self-hosted results may differ.

Details: Apache-2.0 · Python · mem0ai/mem0

2. MemPalace

Best for: memory that stays on your machine.

MemPalace stores conversations verbatim, supports pluggable backends and runs with no API calls. It was created in April 2026 and has grown very fast.

Why it made the list: local-first memory matters when the data is private.

Heads-up: its 96.6% score is retrieval recall at five results on LongMemEval, reported by the project itself.

Details: MIT · Python · MemPalace/mempalace

3. Hindsight

Best for: agents that should improve, not just recall.

Hindsight says most memory systems focus on recalling history and that it focuses on learning from it. It also offers a hosted cloud version.

Why it made the list: it takes a different approach from plain retrieval.

Heads-up: it is a newer project, created in October 2025, so expect faster change.

Details: MIT · Python · vectorize-io/hindsight

4. Cognee

Best for: turning documents, code and conversations into a knowledge graph.

Cognee builds a self-hosted knowledge graph your agents can search. It says it runs locally for free using small models, with no API key.

Why it made the list: graph memory suits questions that connect many facts.

Heads-up: graph building adds setup and compute compared with a simple vector store.

Details: Apache-2.0 · Python · topoteretes/cognee

5. Graphiti

Best for: facts that change, such as customer details or project status.

Graphiti, from the Zep team, builds temporal context graphs. Unlike a static graph, it tracks how facts change over time.

Why it made the list: knowing what was true and when avoids stale answers.

Heads-up: it is a framework, so you wire it into your own agent and graph database.

Details: Apache-2.0 · Python · getzep/graphiti

6. Supermemory

Best for: a fast memory API with a self-hosting option.

Supermemory calls itself a memory and context engine and can run fully locally. It also offers a hosted dashboard.

Why it made the list: it is popular and actively updated, with strong TypeScript tooling.

Heads-up: its claim of ranking first on major benchmarks is self-reported.

Details: MIT · TypeScript · supermemoryai/supermemory

7. agentmemory

Best for: coding agents that forget your project between sessions.

agentmemory adds persistent memory to Claude Code, GitHub Copilot CLI, Cursor, Gemini CLI, Codex CLI and any MCP client. It was created in February 2026.

Why it made the list: it targets a daily pain point for developers.

Heads-up: it is young, so check how it stores data before you point it at private code.

Details: Apache-2.0 · TypeScript · rohitg00/agentmemory

8. Letta

Best for: building agents whose memory is part of the runtime.

Letta was formerly MemGPT and describes itself as a platform for stateful agents. Its README says the current source lives in the letta-code repository.

Why it made the list: it is one of the original memory-first agent projects.

Heads-up: it is a whole agent platform, so it is more than a drop-in memory library.

Details: Apache-2.0 · letta-ai/letta

9. OpenViking

Best for: organising memory, knowledge and skills in one place.

OpenViking, from Volcengine, is a context database that arranges everything an agent knows as a virtual filesystem. It aims to make stored context visible rather than a black box.

Why it made the list: the filesystem idea is easy to inspect and debug.

Heads-up: it uses the AGPL-3.0 licence, which can require sharing changes to software you offer over a network.

Details: AGPL-3.0 · Python · volcengine/OpenViking

10. EverOS

Best for: one memory layer you own across many tools.

EverOS stores conversations, files and agent history as readable Markdown. It keeps local SQLite and LanceDB indexes for fast retrieval.

Why it made the list: plain Markdown memory is portable and easy to audit.

Heads-up: it is a Python library and runtime, so you integrate it yourself.

Details: Apache-2.0 · Python · EverMind-AI/EverOS

Honorable mentions

These projects just missed the top ten.

• MemOS (12.2K, Apache-2.0): a memory operating system for agents.
• Memori (17.3K): enterprise-focused memory. GitHub could not detect its licence, so read the licence file first.
• Basic Memory (4.5K, AGPL-3.0): Markdown notes your AI can read and write through MCP.
• LangMem (1.7K, MIT): memory tools for LangGraph users.

The infrastructure underneath: vector databases

Most memory layers store embeddings in a vector database. These are the largest open-source options.

If you already run Postgres, pgvector can be the simplest start. Our PostgreSQL skill covers setup and tuning.

Benchmarks: what the numbers do and do not tell you

Several projects publish scores on LoCoMo and LongMemEval. They are measured differently, so you cannot compare them directly.

• mem0 reports answer accuracy on its managed platform.
• MemPalace reports retrieval recall at five results.
• Supermemory reports recall at fifteen results and rankings it ran itself.

All of these are vendor-reported. Test on your own data before you trust a ranking.

How to test a memory layer yourself

A one-day test beats any leaderboard.

1. Write 50 questions from your real conversations or documents.
2. Measure answer accuracy, tokens per query and latency.
3. Change a fact and check the agent uses the new one.
4. Delete a record and confirm it is really gone.
5. Check that one user's memory never appears for another.

Licences matter for infrastructure

Memory sits at the centre of your product, so the licence matters. Apache-2.0 and MIT are permissive.

AGPL-3.0 can require you to share source for modified versions you run as a service. If a project's licence is unclear, ask a lawyer before you build on it.

Security and privacy for agent memory

Memory stores personal and business information, so treat it as sensitive data. Limit what is stored, encrypt it and give users a way to delete it.

If you expose memory through an MCP server, review its access carefully. Our skills can help: Agent Memory MCP, MCP Builder and MCP Server Security.

Browse more in the AI Agents & MCP category.

Which memory repo should you choose?

• General app with a managed option: mem0 or Supermemory.
• Private, local-only data: MemPalace, EverOS or Cognee.
• Facts that change over time: Graphiti.
• Coding agents: agentmemory.
• A full stateful agent platform: Letta.
• Agents that learn from outcomes: Hindsight.

Frequently asked questions

What is AI agent memory?

It is a layer that stores and retrieves information between sessions. It lets an agent remember users, decisions and past work.

What is the difference between memory and RAG?

RAG retrieves from a fixed set of documents. Memory also writes new information as the agent works, and it updates or forgets it.

What are the best mem0 alternatives?

MemPalace, Cognee, Graphiti and Supermemory are the closest open-source options. Pick by whether you need local storage, graphs or time-aware facts.

Do I need a vector database for agent memory?

Often yes, but not always. Some projects use graphs or Markdown files, and Postgres with pgvector is enough for many apps.

Are these projects free?

The code is free to use under its licence. Several also sell hosted versions, which cost money.

Methodology and disclosure

GitHub data was pulled from the GitHub API on November 2, 2026, and project details come from each README. Stars and benchmark claims change quickly.

We have not run these projects. This article contains no affiliate links.