Milla Jovovich Builds MemPalace AI Memory System with Claude
Hollywood Star Meets AI Engineering
In a groundbreaking fusion of entertainment and engineering, Milla Jovovich — acclaimed actress of The Fifth Element and Resident Evil — co-created MemPalace, an open-source AI memory system built with Anthropic’s Claude. Developed alongside engineer Ben Sigman, the project achieved a historic perfect score (500/500) on the LongMemEval benchmark — the first-ever full-score result for long-term memory evaluation.
Architectural Innovation: The Memory Palace Framework
Inspired by the ancient Greek Method of Loci, MemPalace structures knowledge in a hierarchical, navigable virtual space:
- Wings: Represent projects, people, or domains (e.g.,
Driftwood,Kai). - Rooms: Thematic subspaces (e.g.,
auth migration,database selection). - Halls: Fixed memory-type corridors (
decisions,milestones,preferences,suggestions,discoveries). - Tunnels: Auto-generated cross-wing links for shared concepts.
- Closets & Drawers: Summarized indexes + verbatim full transcripts — zero loss, full fidelity.
This architecture enables precision retrieval: search scope narrows from full corpus → wing → room → drawer. On 22,000+ real dialogues, recall jumped from 60.9% to 94.8% — a +34 percentage point gain — simply via metadata filtering.
GitHub Repository: https://github.com/milla-jovovich/mempalace
AAAK: AI-Optimized Structured Notation
MemPalace introduces AAAK — a human-readable yet AI-native shorthand language designed for models, not people:
TEAM: PRI(lead) | KAI(backend,3yr) SOR(frontend) MAY(infra) LEO(junior,new)
PROJ: DRIFTWOOD(saas.analytics) | SPRINT: auth.migration→clerk
DECISION: KAI.rec:clerk>auth0(pricing+dx) | ★★★★
- Input: ~1000 tokens (natural English)
- AAAK encoding: ~120 tokens (8× compression, lossless semantics)
- Universally parsable by Claude, GPT, Gemini — no decoder or fine-tuning required.
Transparent Iteration & Community Audit
Within 48 hours of launch, open-source contributors conducted rigorous scrutiny:
| Claim | Reality | Resolution |
|---|---|---|
| “30× lossless compression” (AAAK) | Token count increased in some cases; actually lossy | Public correction: renamed “lossy structured shorthand” |
| “+34% palace gain” | Metadata filtering is ChromaDB-native — not novel logic | Clarified as standard capability, not proprietary advantage |
| “Auto fact-checking via KG” | fact_checker.py was standalone, unconnected |
Patch tracked in GitHub Issue; integration planned |
The team published a candid public letter acknowledging every discrepancy — headlined: “We’d rather be right than look impressive.”
Cost-Efficiency & Local-First Design
- Zero API calls: All data stored locally in ChromaDB.
- No cloud fees: Runs entirely offline.
- $0.70/year for heavy users (~19.5M tokens/year), versus $507/year for pure LLM summarization.
- Retrieval loads only ~170 tokens of critical context per session — dynamic, lean, precise.
Quick Start Guide
pip install mempalace
# Initialize your memory world
mempalace init ~/projects/myapp
# Mine diverse sources
mempalace mine ~/projects/myapp # code, docs, notes
mempalace mine ~/chats/ --mode convos # Slack, ChatGPT, Claude exports
mempalace mine ~/chats/ --mode convos --extract general # classify decisions/milestones
# Search instantly
mempalace search "why did we switch to GraphQL"
# Connect to MCP-compatible tools (Claude, Cursor, etc.)
claude mcp add mempalace -- python -m mempalace.mcp_server
Cultural Significance: Redefining the Developer Identity
MemPalace transcends technical novelty. It signals a paradigm shift: the boundary between creator and developer is dissolving. A globally recognized film icon — leveraging Claude as a collaborative coding partner — co-led a project that outperformed industry giants on a key SOTA benchmark. And it’s open, free, local, and production-ready.
Ben Sigman’s playful naming nod — MemPalace → Multipass — echoes Leeloo’s iconic line from The Fifth Element: “Multipass.” This time, the pass truly is unrestricted.
References:
– @bensig on X
– @wangray on X
– GitHub: milla-jovovich/mempalaceArticle originally published by XinZhiYuan; author: “Peach — So Tired”.