claude-mem gives AI coding agents persistent memory across sessions. It runs as a plugin that hooks into the agent lifecycle: as you work, it captures tool-usage observations, generates semantic summaries of each session, and stores them so that later sessions can automatically reload relevant context. The goal is continuity, so an agent retains knowledge about a project after a session ends or reconnects, without manual note-keeping. It is built primarily for Claude Code but also supports OpenCode, Antigravity CLI, and OpenClaw gateways, among other environments.
Under the hood it combines a local worker service (an HTTP API with a web viewer UI, managed by Bun), a SQLite database holding sessions, observations, and summaries, and a Chroma vector database for hybrid semantic-plus-keyword retrieval. Memory is queried through a mem-search skill and a set of MCP tools organized as a three-layer, token-efficient workflow: search returns a compact index of matching IDs, timeline gives chronological context around a result, and get_observations fetches full details only for the IDs you select, which the project reports as roughly a 10x token saving over fetching everything.
Additional features include progressive-disclosure retrieval with visible token costs, per-project search, configurable workflow modes and output languages (for example code, code--zh, code--ja), and privacy controls via tags to exclude sensitive content from storage. Note that a separately promoted CMEM crypto token mentioned in the README is a third-party creation, not part of the tool itself.
