MemPalace

MemPalace might be the most unlikely collaboration in open source this year: Hollywood actress Milla Jovovich (Resident Evil, The Fifth Element) and Bitcoin entrepreneur Ben Sigman spent months building an AI memory system with Claude — with Claude Opus 4.6 listed as a co-author in the git history. The repository now has over 59,000 stars and was still being pushed to today.

The core idea is borrowed from the ancient Greeks: the method of loci, or memory palace. Instead of letting an AI decide what to forget via summarization, MemPalace stores conversations verbatim and organizes them into a navigable hierarchy — wings, halls, rooms, closets, drawers, tunnels — backed by ChromaDB for semantic search and SQLite for a temporal knowledge graph. It runs entirely on your machine, exposes 19 MCP tools so Claude Code, Cursor or ChatGPT can read and write memories mid-session, and ships under MIT.

Then came the benchmark drama, which is honestly the most instructive part. The team claimed 96.6% on LongMemEval with zero API calls, plus a first-ever perfect 100% R@5 with Haiku reranking — and the internet promptly audited them. Critics found methodology issues; the authors published a correction note admitting the 96.6% was raw verbatim mode and that their experimental compression dialect actually regressed performance. The 96.6% raw-mode score has been independently reproduced, so the core result stands, but the episode is a useful reminder.

MemPalace earns its place here for two reasons. First, it's a serious piece of AI infrastructure that was genuinely built with AI assistance — the snake eating its own tail. Second, the benchmark controversy played out in public, with unusual honesty from the authors. That's the healthy version of vibe-coding culture: ship fast, claim loudly, but publish your eval code and correct the record when you're wrong.

Note: This project is curated by the VibeFix editors from public sources. All rights belong to the original creator.