// the find
MemPalace/mempalace
The best-benchmarked open-source AI memory system. And it's free.
MemPalace is a local-first memory layer for AI assistants that stores conversation history verbatim and retrieves it via semantic search, with no summarization or paraphrasing. It structures memory hierarchically — people and projects become wings, topics become rooms, content lives in drawers — so queries can be scoped rather than run against a flat blob. Aimed at developers who want persistent memory across Claude Code, Cursor, or any MCP-compatible tool without sending data to a cloud service.
The benchmark transparency is rare and genuinely good: raw result files are committed to the repo, methodology is in BENCHMARKS.md, and they explicitly flag where the ≥99% reranked number is 'teaching to the test.' Most AI memory projects paste a headline number with no reproducibility path. The verbatim storage principle is well-reasoned — no summarization means no lossy compression of history, and scoped retrieval (by wing or room) avoids brute-forcing a flat corpus every time. The pluggable backend abstraction holds up: five backends (ChromaDB, SQLite, Milvus, Qdrant, pgvector) with a shared conformance suite in tests/_backend_conformance.py, and the README explicitly notes they chose 'deliberately different substrates' to avoid accidentally shaping the interface around one vendor. The zero-cloud core path is a genuine differentiator — 96.6% R@5 on LongMemEval with no API key and no outbound calls, using a local embedding model.
ChromaDB as the default backend is the weakest architectural choice: gRPC overhead, memory pressure on large collections, and a history of breaking releases between minor versions mean users will eventually hit a wall without realizing they need to swap backends. The self-declared 'best-benchmarked' in the repo description is a marketing claim that the repo itself undercuts — they refuse head-to-head comparisons with Mem0 or Zep (citing metric incompatibility, which is fair), but then claim the superlative anyway. The codebase has grown idiosyncratic fast: wal.py, dialect.py, hallways.py, closet_llm.py, write_routing.py — the abstractions are creative but the naming convention is its own dialect, and contributing or debugging requires learning a non-standard mental model before touching anything. The prominent CAUTION block about impostor sites and malware indicates the project has already had or is actively managing security incidents around its brand, which is worth noting for anyone considering it in a team or production context.