# cavemem

cavemem gives coding assistants cross-agent persistent memory: hooks capture session observations, compress them into a local SQLite store, and expose them back through MCP tools. It is now frozen; its memory core continues in a successor, caveman.

cavemem gives coding assistants persistent memory that carries over between sessions and across different agents. Hooks fire at session boundaries and compress observations using a deterministic caveman grammar that keeps code and paths byte-for-byte while cutting prose token count by about 75%, then write the result to a local SQLite database with no network calls.

Agents retrieve memory progressively through three MCP tools, search, timeline and get_observations, using hybrid SQLite full-text and local vector search. It supports capturing observations from Claude Code, OpenCode, Codex, GitHub Copilot and Augment Code, and query-only MCP search from Cursor, Gemini CLI, Antigravity and IBM Bob. A local, token-protected web viewer shows past sessions in a human-readable form, and path globs or explicit markers can exclude sensitive content from being written.

The project is frozen as of August 2026 and receives no new features or fixes, though it still installs and works; its compressed-memory core continues in a successor project, caveman.

Website: https://github.com/JuliusBrussee/cavemem