Your coding assistant's goldfish memory, now with rocks

What it does

cavemem hooks into coding assistants at session boundaries, strips private content, compresses observations with a deterministic “caveman grammar,” and drops everything into local SQLite. Agents query back through three MCP tools — search, timeline, get_observations — with hybrid BM25 + vector ranking. No cloud, no daemon to babysit. A background worker auto-spawns for embeddings and quits when bored.

The interesting bit

The compression is the product. Prose gets squashed to ~75% fewer tokens (“auth mw throws 401 @ session token expires”), but code blocks, paths, and identifiers stay byte-for-byte intact. Round-trip expansion is guaranteed — the stored gibberish rehydrates to readable text for the web viewer at localhost:37777. It’s a deliberately lossy pipe that knows exactly where not to leak.

Key highlights

• Cross-IDE installers for Claude Code, Cursor, Gemini CLI, OpenCode, and Codex — one command each
• Progressive MCP retrieval: compact summaries first, full bodies only on demand via get_observations
• Hybrid search with tunable BM25/vector blend (search.alpha)
• Privacy boundaries: <private> tags stripped at write time, path globs exclude whole directories
• Optional remote embedding providers (Ollama, OpenAI) if local isn’t enough
• Part of a four-repo “Caveman Ecosystem” — output compression, memory, build loops, and a fine-tuned Gemma model

Caveats

• Hook handlers target sub-150ms completion; heavy sessions may stress that budget
• The “caveman grammar” compression is proprietary and not documented beyond examples — you’ll trust the round-trip guarantee or you won’t
• Ecosystem cross-selling is aggressive; each repo stands alone, but the README really wants you to collect all four

Verdict

Worth a look if you bounce between Claude Code and Cursor and want one memory store that follows you. Skip it if you already have a RAG pipeline you like, or if “deterministic gibberish rehydration” sounds like a bug report waiting to happen.