Milla Jovovich Co-Creates MemPalace, an Open-Source AI Memory System That Scored 96.6% on LongMemEval

Actress Milla Jovovich, best known for her roles in Resident Evil and The Fifth Element, has co-developed an open-source AI memory system called MemPalace alongside developer Ben Sigman. The project achieved 96.6% recall on LongMemEval without any API calls — the highest score recorded for a fully local memory system.

Key Highlights

• 96.6% R@5 on LongMemEval with zero API calls; the team later reported 100% with Haiku reranking, though this claim has been disputed (see caveats below)
• Fully local and free: runs on SQLite and ChromaDB with no cloud services, no subscriptions, and no external APIs required
• MIT licensed and available on GitHub, with Python 3.9+ support and MCP integration for Claude, ChatGPT, and Cursor

How MemPalace Works

MemPalace applies the ancient Greek method of loci — the memory palace technique — to AI agent memory. Instead of letting AI decide what to remember, it stores every word and organizes information into a structured hierarchy:

• Wings represent projects or people
• Halls categorize memory types such as facts, events, and preferences
• Rooms hold specific topics within each hall
• Closets contain compressed summaries pointing to original content
• Tunnels create cross-references connecting rooms across different wings

This architectural approach dramatically improves retrieval accuracy. Searching within a wing boosts accuracy from 60.9% to 73.1%, while adding hall and room context pushes it to 94.8% — a 34% improvement over flat, unstructured search.

The AAAK Compression Dialect

One of MemPalace's most innovative features is AAAK, a shorthand dialect designed for AI agents that achieves roughly 30x compression. A passage that would normally consume around 1,000 tokens compresses to approximately 120 tokens, enabling AI to load months of context before the first message.

AAAK works natively with any large language model — Claude, GPT, Gemini, Llama, and Mistral — without requiring specialized decoders or fine-tuning.

Why It Matters

AI assistants today lose all context when a session ends. Months of decisions, debugging sessions, and architectural discussions simply vanish. MemPalace addresses this by preserving everything locally, making it searchable and retrievable across sessions.

The system was tested on more than 22,000 real conversation memories and also scored 92.9% on ConvoMem — more than double the score of Mem0, one of the leading paid memory solutions.

Additional Features

Beyond basic memory storage, MemPalace includes a knowledge graph with temporal entity relationships, contradiction detection across facts, specialist agent support with individual diaries, and three mining modes for projects, conversations, and general classification. It exposes 19 MCP tools for integration with popular AI coding environments.

Benchmark Caveats

Community reviewers have raised concerns about MemPalace's benchmark methodology:

• The 100% LongMemEval claim involved post-hoc tuning. The team identified the 3 specific questions the system missed and engineered targeted fixes (quoted-phrase boosts, person-name boosts) before retesting on the same set. The pre-tuning score of 96.6% is the more defensible number.
• The 100% LoCoMo score is inflated by retrieval window size. LoCoMo sessions contain 19–32 items, but MemPalace ran with top_k=50 — a retrieval window larger than the candidate pool, meaning nearly everything is retrieved by definition.
• The 96.6% raw score uses ChromaDB default embeddings on uncompressed text. The palace hierarchy (wings, rooms, halls) is not exercised during benchmarking.

These caveats do not invalidate the project's architectural innovations, but readers should weigh the headline numbers accordingly.

What Is Next

MemPalace v3.0.0 was released on April 5, 2026, and has already attracted attention on Hacker News and across social media. The project has an active Discord community and welcomes contributions. Its MIT license and fully local architecture make it particularly appealing for developers concerned about privacy and vendor lock-in.

Source: MemPalace on GitHub