You know Milla Jovovich, right? The one who plays Alice in the *Resident Evil* movies.

A few days ago, she open-sourced an AI memory system on GitHub called MemPalace. It achieved the first perfect score on an industry-standard benchmark—96.6%, and with Haiku reranking, it can reach 100%.

What's the story? Reportedly, after months of using AI conversations, she lost a vast amount of her accumulated decisions and thought processes. Existing memory systems let the AI decide what's worth remembering, which wasn't what she wanted.

So, she and a friend spent several months building MemPalace using Claude Code. The core idea borrows from the ancient Greek 'Method of Loci' or memory palace technique—organizing memories into a navigable spatial structure (Wings, Rooms, Closets, Drawers) rather than letting the AI filter what it should remember. This structure alone resulted in a 34% retrieval improvement.

There's also a compression protocol called AAAK, offering 30x lossless compression. What does that mean? Typically, six months of AI conversations amount to about 19.5 million tokens. With MemPalace, you only need about 170 tokens each time to "wake up" the AI and make it remember your world. When you need to search, it adds around 13,500 tokens. The estimated annual cost is roughly $10.

It runs purely locally, is MIT licensed, and requires no APIs.

However, a commenter clarified: She is not the developer. The main developer is Ben Sigman (benobi.one). Milla was more like the architect, providing the ideas and vision, which she has acknowledged on Twitter herself. In modern terms, this is a classic "vision provider + technical implementer" founder pairing.

Interesting.

---

**Repository Address**: https://github.com/milla-jovovich/mempalace

**Technical Details** (from GitHub):

- LongMemEval R@5 score: 96.6% (zero API calls)

- Can reach 100% with Haiku rerank

- AAAK Compression: 30x lossless, compatible with any LLM

- Knowledge Graph: Temporal entity relationships, SQLite local storage

- MCP Protocol: 19 tools, compatible with Claude/Cursor/GPT

- Four-layer Memory Stack: L0 Identity (50 tokens) + L1 Key Facts (120 tokens) + L2 Load-on-Demand + L3 Deep Search

An actress has created something that surpasses all existing AI company memory products—truly an "age of imagination."