Verbatim AI memory that runs offline and benchmarks honestly

What it does

MemPalace indexes your conversation history and project files as verbatim text, then retrieves them with semantic search. It organizes everything into a structured hierarchy—people and projects become wings, topics become rooms, and raw content lives in drawers—so searches can be scoped instead of drowning in a flat corpus. The default vector backend is ChromaDB, but the retrieval interface is pluggable if you want to swap it out.

The interesting bit

The project refuses to summarize or paraphrase your data, betting that raw text plus smart retrieval beats compressed lossy memory. That retrieval works entirely offline—no API keys, no cloud, no LLM required—and still hits 96.6% R@5 on the LongMemEval benchmark, a figure they document down to the per-question result files.

Key highlights

• Stores text verbatim without summarization, extraction, or paraphrasing
• Core semantic search runs fully offline with a local embedding model (~300 MB) and ChromaDB; no API key needed
• Reproducible benchmarks: 96.6% R@5 raw, 98.4% hybrid held-out on LongMemEval; 88.9% R@10 on LoCoMo
• Structured storage with wings, rooms, and drawers for scoped search
• Built-in temporal knowledge graph (validity windows, SQLite-backed) and 29 MCP tools

Caveats

• The authors warn of active impostor sites using similar domains that may distribute malware; they publish a detailed timeline in docs/HISTORY.md
• Claude Code sessions expire in 30 days unless you wire the auto-save hooks, which is presented as an urgent setup step rather than an afterthought

Verdict

Worth a look if you want AI assistants that remember past conversations without shipping your logs to a vendor. Skip it if you need a managed cloud memory service or prefer systems that compress history into summaries.