compare
recall vs claude-mem
claude-mem is a popular open-source memory for coding agents. recall is decision memory for a team. They do different jobs; this page describes both, with sources.
what it is
What claude-mem is.
In its own words, a "persistent memory compression system built for Claude Code".
It is open source under the Apache 2.0 license. Its GitHub repository showed about 94,800 stars on September 27, 2026.
It captures through lifecycle hooks: session start, prompt submit, after each tool use, stop and session end. A worker built on the Claude Agent SDK asks a model to compress tool-use observations into typed observations and session summaries.
It stores them on your machine in SQLite with full-text search, with optional Chroma vector search, behind a local worker that also serves a web viewer.
A new session gets context in layers: a compact index first, then a timeline, then full observations only for what is needed.
Text wrapped in <private> tags is used in the session and not stored.
Its repository lists Claude Code, Codex, Gemini, Copilot, OpenCode and more, and its docs cover Cursor.
Sources, read September 27, 2026: github.com/thedotmack/claude-mem (README and license) and docs.claude-mem.ai (hooks, database, search, worker service, private tags).
what it does well
What claude-mem does well.
For one person on one machine, it is a strong default.
- Local and open source: the database lives on your machine.
- Automatic: install it and sessions are captured with no action from you.
- Careful with tokens: layered retrieval brings back full detail only when it is needed.
- Easy to inspect: a web viewer shows the memory stream as it is written.
we read its source in September 2026 and borrowed one idea: analytics shows characters read in against characters handed out.
where they differ
Where they differ.
Only on the three mechanics, provenance and permissions. We have run no benchmark against claude-mem, so there is none here.
| Question | claude-mem, from its published docs | recall |
|---|---|---|
| Two notes disagree | No contradiction handling found | A person settles it; a model never erases either |
| Did a decision hold? | A decision observation type exists, as unstructured text; no outcome field found | Held, reversed or regretted; reversed ones are flagged in every later brief |
| Who may read and write what? | One person, one local database | Spaces, grants and agent ceilings; no agent can write a rule that binds a person |
| Who said it? | Your prompts are stored apart from observations; no per-record speaker field found | Every record says a person or the agent, with the quote it rests on |
| Why, and when does it hold? | Unstructured narrative and facts | because and holds when on every record |
| Where it runs | Your machine | A service at one MCP endpoint, early access by invitation |
which fits
Which one fits.
claude-mem if you are one person who wants coding sessions remembered on your own machine, at no cost. recall if a team shares decisions across people and agents and needs to show who decided what, and when.
01Have you benchmarked recall against claude-mem?
No. The repeat-rate test, a CLAUDE.md file vs claude-mem vs recall on the same corrected mistakes, is not measured yet. It will be published on /trust with its denominator.
02Is anything on this page from claude-mem's roadmap?
No. It describes only what its README and published docs said on September 27, 2026. If that changes, this page should change too.
03Where does my data live with recall?
In the workspace's spaces on the recall service. Extraction runs on the model endpoint you configure, including a local one, and any space exports as Markdown or JSON.
early access
Connect, once you have access.
Early access is by invitation while we load-test capacity and cost. With access, every MCP client uses one endpoint.
https://app.recall.hiretrainva.com/mcp