claude-mem is a persistent memory system for Claude Code that captures what the agent does in each session, compresses it with AI and offers it back in later sessions. On the repo page, checked on October 14, 2026, it shows about 100.5k stars and 8.8k forks, and GitHub's weekly trending list credits it with about 3,587 new stars. It is a noisy repo right now, so it is worth looking at what it actually does.
What it is
The problem is a familiar one: every coding-agent session starts without remembering the last. claude-mem stores observations of tool use, generates summaries and makes them searchable. Per the README it supports Claude Code, Codex, Gemini, Copilot, OpenCode and others.
It works through six lifecycle hooks (SessionStart, UserPromptSubmit, PreToolUse, PostToolUse, Stop and SessionEnd), each run as claude-mem hook claude-code <event>. The hooks talk to a local service with an HTTP API and web viewer, managed with Bun. Storage is SQLite for sessions, observations and summaries, plus a Chroma vector database for hybrid semantic and keyword search. The stack is TypeScript on Node.js 20+, with Bun and uv.
To query the memory there is a search skill and MCP tools (search, timeline and get_observations), meant to be used in that order: search first, then the timeline, then fetch details for the IDs you care about.
How to use it
Per the README there are two install paths:
# Installer (offers sign-in and provider choice)
npx claude-mem installOr from inside Claude Code, via the plugin marketplace:
/plugin marketplace add thedotmack/claude-mem
/plugin install claude-memThen restart. The README warns that npm install -g claude-mem installs only the SDK, without hooks or the service. Once installed, context from earlier sessions shows up on its own; to search by hand, the suggested pattern is search(query="authentication bug", type="bugfix", limit=10), then timeline, then get_observations(ids=[...]). Anything wrapped in <private> tags is not stored.
What I like and what I don't
I like that storage is local (SQLite and Chroma), that there is an explicit tag to exclude content, and that the license is Apache-2.0, which is clear for professional use. The hook-based design also uses Claude Code's standard mechanism rather than a hack.
What holds me back is privacy. Observations are compressed with AI, so a model provider is involved: the README mentions a hosted observer with a 30-day free trial, or your own OpenRouter or Gemini key, or your Anthropic plan. There is also optional cloud sync (cmem.ai). The page does not detail which data leaves the machine in each mode, and an agent working on production code sees secrets, paths and logs. Before using it anywhere serious I would read the hook code and choose a provider explicitly, or turn off online mode with CLAUDE_MEM_ONLINE_OPTIN=false.
I also could not confirm the date of the last commit or release: the page shows no releases (the badge says version 13.39.0) and lists 34 open issues and 57 pull requests. And I have not installed or run it. The closest alternatives, judging by its topic tags, are mem0, openmemory and supermemory.
Sources
- claude-mem repository on GitHub (README and metadata checked on October 14, 2026).
- GitHub Trending, weekly (October 14, 2026).