Key takeaways
- Your coding agent stops starting from zero: Claude-Mem carries earlier work into later sessions.
- Search is layered so the agent pulls short results first and full detail only when needed.
- The project claims about 10x token savings. That is its own figure; we did not measure it.
- Telemetry is on by default, and open issue #4605 says privacy statements contradict the shipped defaults.
What it is
Claude-Mem is an open-source project that captures what a coding agent does during a session, compresses it with AI and injects relevant context into later sessions. It is licensed under Apache-2.0 and lives at github.com/thedotmack/claude-mem.
The problem it solves
A coding agent forgets everything when a session ends. Anyone working on a long project must re-explain the architecture, earlier decisions and why a fix was made, every time they start again.
What you get
- Continuity: context from earlier sessions is injected when a new one starts.
- Lower token use: search returns short results of about 50 to 100 tokens each, and full detail of about 500 to 1,000 tokens is fetched only for the items you need.
- Visibility: a local web viewer shows the memory stream as it grows.
- Control: content inside private tags is visible in the session but not saved.
- Reach: it works with Claude Code, Codex, Gemini, OpenCode and other agents.
How it compares
The project's README claims about 10x token savings from its three-layer search, which runs search, then timeline, then full observations. That is the project's own claim, and we did not measure it.
The catches
- Telemetry is on by default (opt-out). The project documents how to switch it off.
- Open issue #4605 says privacy statements contradict the shipped defaults. A contributor commented that part of it is now addressed, while the marketplace consent gap appears unchanged.
- Compression uses an AI model, so it consumes usage. The default provider is Claude with the Haiku model.
- The README says CMEM is a token created by a third party. It is unrelated to using the tool.
- A critical command injection reported in December 2025 is listed as resolved.
What we did not test
We did not run Claude-Mem on a real machine, because installing it changes the host agent's configuration. We did not measure the token savings, the quality of the compressed memory or what telemetry sends. Every claim here is tied to a literal quote from the project's own repository.
Verdict
Editors' view: the idea is strong for long projects that span many sessions, and the layered search is a sensible way to keep token use down. But decide on telemetry and privacy, not on the idea alone. If you work on sensitive code, read issue #4605 and the telemetry page first.
Sources
Need help adopting this, or have a software project in mind?
Tell us what you are building. The initial consultation is free. Email [email protected] or follow @zolnaltd.
Start a conversation →