claude-mem

claude-mem is a Claude Code plugin that carries context across sessions. A terminal session ends and everything you explained about the project goes with it. The next morning you start over: what the stack is, what you did yesterday, why that file must not be touched. This plugin closes exactly that gap.

How it works. It attaches to Claude Code at five lifecycle points: session start, prompt submitted, tool call, stop, and session end. At those points it records observations — which files were opened, what got fixed, what you decided. At the end of a session it compresses them into short summaries, and at the start of the next one it feeds back the ones that are relevant.

What lands on your machine. A local HTTP service with a web interface for browsing the memory, a SQLite database as an ordinary file on disk, and a Chroma vector database for meaning-based search — which is what lets you find "that bit where we fixed authentication" without remembering the exact wording. A separate skill lets you question the memory in plain language. No cloud: everything stays local.

The search is built economically, and it is a rare case of token cost being thought through in advance. Instead of dumping everything it finds into the context, it works in three layers: a compact list of results first, then a timeline around whatever looks interesting, and only then the full detail on the entries you picked.

What to know before installing. The familiar `npm install -g claude-mem` completes without errors but installs only the library: the hooks are never registered and the memory never works. Install through `npx claude-mem install` or from the plugin marketplace. You need Node.js 20 or newer; the plugin pulls in Bun and uv itself. The licence is Apache-2.0, so it can go into anything, commercial products included. One privacy note: observations are written to disk in the clear, which is worth weighing if you work with someone else's credentials or with personal data.