Local-first AI memory. Verbatim storage, pluggable backend, and 96.6% R@5 raw recall on LongMemEval — with zero API calls. Open-source and free to use.
Capabilities
A memory architecture designed from first principles — not an afterthought bolted onto existing systems.
Conversations are stored word-for-word — not summarized, extracted, or paraphrased. 96.6% R@5 raw on LongMemEval, 98.4% with the hybrid pipeline.
People and projects become wings, topics become rooms, content lives in drawers — so searches can be scoped, not run against a flat corpus.
A local SQLite-backed entity-relationship graph with validity windows. Add, query, invalidate, and timeline facts as they evolve.
Nothing leaves your machine unless you opt in. No API keys, no cloud, no LLM required for retrieval. MIT-licensed and fully inspectable.
ChromaDB by default, but the retrieval interface is clean — swap in an alternative backend without touching the rest of the system.
Palace reads/writes, knowledge-graph operations, cross-wing navigation, drawer management, and agent diaries — ready for Claude Code, Gemini CLI, and any MCP-compatible tool.
Under the Hood
MemPalace uses a layered memory architecture inspired by human cognition — separating working memory, episodic memory, and long-term semantic storage into distinct but interconnected systems.
Performance
Every number below is reproducible from the public repo. Full per-question result files are committed under benchmarks/results_*.
| Benchmark | Metric | Score | Notes |
|---|---|---|---|
| LongMemEval — raw | R@5 | 96.6% | 500 questions • no LLM, no heuristics |
| LongMemEval — hybrid v4 | R@5 | 98.4% | 450q held-out (tuned on 50 dev) |
| LongMemEval — hybrid v4 + rerank | R@5 | ≥99% | Any capable LLM as reader |
| ConvoMem | Avg recall | 92.9% | 250 items, 50 per category |
| LoCoMo — hybrid v5 | R@10 | 88.9% | 1,986 questions, no rerank |
| MemBench (ACL 2025) | R@5 | 80.3% | 8,500 items, all categories |
MemPalace deliberately does not headline a “100%” number: the last 0.6% was reached by inspecting specific wrong answers, which the project flags as teaching to the test. 96.6% R@5 raw and 98.4% hybrid held-out are the honest, generalisable figures.
The Team
Architect
Actress, musician, model — and the architectural vision behind MemPalace. Designing how AI should remember.
Engineer
Building the core engine, benchmarks, and infrastructure that makes perfect recall a reality.
Get Started
MemPalace is free, open-source, and ready to integrate. Whether you're building a chatbot, an agent, or something entirely new — give it the memory it deserves.