MemPalace — Local-First AI Memory CLI

Practical Notes

• Quant: LongMemEval raw retrieval recall is 96.6% R@5 (500 questions) without any LLM calls.
• Quant: the README also reports 98.4% R@5 on a held-out 450-question split (Hybrid v4) with no LLM required.

Main

Use MemPalace like a “memory pipeline”, not a chat add-on:

1. Mine two sources: your repo + your agent transcripts (Claude Code sessions).
2. Search before you re-explain: ask MemPalace for the decision history instead of retyping context.
3. Wake up at session start: run mempalace wake-up and paste only the returned context into the new chat.

Watchouts

• The project warns about impostor domains; treat installers outside GitHub/PyPI/docs as untrusted.
• Memory can still leak secrets if you mine private folders into a shared store—scope mines per project.

FAQ

Q: Is an LLM required to get value? A: No. The raw retrieval benchmark reported is 96.6% R@5 with no LLM calls.

Q: What should I mine first? A: Start with one repo and one transcript folder, then expand only after retrieval stays high-signal.

Q: How do I keep memory scoped? A: Mine per project path and avoid mixing unrelated repos into one shared store.