agent-memory
agent-memory is an open-source, CLI-based long-term memory runtime that stores AI agent memories as Markdown files, for developers using Claude Code or Codex.
No visits from OpenBeat in the last 30 days yet. Every click on “Visit” is counted here.
Pricing and API checked by an OpenBeat editor on Nov 11, 2026 · facts from the maker’s own site ·
Tool insights
agent-memory keeps what an AI agent learns in plain Markdown files that you can read and edit, and searches them locally with ranked retrieval. Several agents, such as Claude Code and Codex, can share one memory store.
AI coding agents forget things between sessions, and memory kept in opaque databases is hard to inspect or fix. agent-memory puts memory in readable files and adds recall, consolidation and history tracking on top.
Developers who run command-line coding agents and want inspectable, local memory shared across tools. People who want a hosted service, a packaged install from PyPI or a polished interface should look elsewhere.
What it does
agent-memory is a long-term memory runtime for AI agents, made by tigerless-labs. Memories live in Markdown files, and those files are the source of truth. A local index built on SQLite FTS5 handles search, so recall works without a hosted service.
The main use case is giving coding agents such as Claude Code and Codex a memory that lasts across sessions. It runs as a command-line tool.
Key features
- Markdown as source of truth: memories are plain files you can open and edit.
- Local recall: BM25 ranking over an SQLite FTS5 index. Recall returns L0 path lists.
- Optional vector plugin: results from vector search can be fused with BM25 using reciprocal rank fusion (RRF).
- Sleep-time consolidation: a "Dreaming" process tidies memory in the background. Deletions need confirmation. It can be started by cron or by hooks.
- History kept on change: when a memory is superseded, the old version stays, and
recall --as-oflets you query an earlier state. - Shared store: Claude Code and Codex can use the same store.
- MCP server: the
mem-mcpcommand exposes memory over MCP. - Separate reasoners: you can choose one reasoner for Write and another for Manage.
Pricing
agent-memory is open source under the MIT license. As of November 2026, the maker's page lists no paid plans and no price. Whether a hosted or free-tier offering exists is not stated on the maker's site.
API and integrations
A public API is not stated on the maker's site, and no SDKs are listed. Listed integrations are Claude Code, Codex CLI, Muse Code, Hermes and MCP. The only listed platform is the CLI. The maker's experiments used Claude Haiku 4.5 as the host and Sonnet 5 as the judge, and a config example references google/gemini-3.7-flash.
Good to know
- There is no PyPI release. You install from a checkout of the repository.
- It needs Python 3.12 or higher and uv.
- Codex recall needs the store to be writable from outside its workspace.
- Raw Session inspection is experimental.
- The native Manage reviewer is experimental.
- No documentation, pricing or changelog page is linked; the GitHub repository is the only listed link.
FAQ
Is agent-memory free and open source?
Yes. agent-memory is open source under the MIT license. As of November 2026, no paid plans or prices are listed on the maker's site.
How do I install agent-memory?
There is no PyPI release, so you install it from a checkout of the GitHub repository. It requires Python 3.12 or higher and uv.
Which AI agents work with agent-memory?
Listed integrations are Claude Code, Codex CLI, Muse Code, Hermes and MCP. Claude Code and Codex can share one memory store. An MCP server is available through mem-mcp.
How does agent-memory search its memories?
It uses local BM25 ranking over an SQLite FTS5 index built from the Markdown files. An optional vector plugin can be added, with results fused by reciprocal rank fusion (RRF).
Does agent-memory keep old versions of a memory?
Yes. When a memory is superseded, the history is kept. The `recall --as-of` option lets you look up what was stored at an earlier point.
Releases
No tracked releases yet.
In the news
No coverage yet.