A Hollywood Actress Built an AI Memory System. It Broke Every Benchmark.

Milla Jovovich & Ben Sigman

Why They Built It — AI Amnesia at Scale

The core frustration was this: six months of daily AI use produces approximately 19.5 million tokens of conversations — every architecture decision, every debugging session, every "we tried X and it failed because Y" — all trapped in chat windows that disappear when the session ends.

Existing tools like Mem0 and Zep tried to solve this by having an LLM extract "key facts" and discard the rest. This sounds efficient but has a fatal flaw: the reasoning — the why behind the decision — is exactly what gets discarded. You're left with "user prefers Postgres" but not the conversation where you explained that it was because your dataset would exceed 10GB and you needed concurrent writes.

"I wanted my AI to remember the way I remember — not just the conclusions, but the journey. The alternatives I considered, the reasons I changed my mind, the nuance. That's what existing memory systems throw away."

— Milla Jovovich, on the motivation behind MemPalace

The Radical Idea: Store Everything, Make It Findable

The fundamental architectural decision in MemPalace is the opposite of every other tool: don't summarise, don't extract, don't decide what matters — store every word verbatim in ChromaDB and use semantic vector search to find it.

The name and structure came from the ancient Greek technique of memorising speeches by mentally placing ideas in rooms of an imaginary building — the "method of loci" or "memory palace." Walk through the building, find the idea. MemPalace applies this exact structure to AI memory:

• Wings — one per person or project (e.g. wing_kai, wing_driftwood)
• Rooms — specific topics within a wing (e.g. auth-migration, graphql-switch)
• Halls — corridors connecting rooms by memory type (facts, events, discoveries, preferences, advice)
• Tunnels — cross-wing connections where the same room topic appears in multiple wings
• Closets — plain-text summaries that point to source content (AAAK-encoded closets coming in a future update)
• Drawers — the original verbatim files, never deleted, never summarised

Built Using Claude Code — Months of Work

MemPalace was built over several months using Claude Code — Anthropic's AI-assisted coding tool. Milla Jovovich described the collaboration: her vision for how AI memory should work, combined with Ben Sigman's engineering expertise to design the architecture, implement ChromaDB storage, write the MCP server, and build the AAAK compression dialect.

The repository (github.com/milla-jovovich/mempalace) is written in Python (98.6%) with Shell scripts (1.4%) and contains:

26,900 Stars in 72 Hours — and the Community's Response

MemPalace v3.0.0 launched on April 6, 2026. Within 24 hours it had accumulated 5,400 GitHub stars; by 72 hours, 26,900. Ben Sigman's launch tweet hit 1.5 million+ impressions. The story caught fire: an A-list actress with a GitHub account had co-built the highest-scoring free AI memory tool ever benchmarked. HackerNews's top comment: "Missed opportunity to call it Resident Eval."

The community also caught real problems in the original README within hours — overstated AAAK compression claims, a misleading token count example, an ambiguous "+34% palace boost" stat. The team responded the same day with a transparent "Note from Milla & Ben — April 7, 2026" that addressed every criticism directly:

• The AAAK token example used a rough heuristic (len(text)//3) instead of a real tokenizer — corrected
• "30x lossless compression" was overstated — AAAK is lossy and regresses LongMemEval from 96.6% to 84.2%
• "+34% palace boost" compares filtered vs unfiltered search, not a novel algorithm — clarified
• "Contradiction detection" wasn't yet wired into the knowledge graph — flagged as in-progress
• "100% with Haiku rerank" is real but the pipeline wasn't in public benchmark scripts — being added

That honesty — rare in the world of AI launch announcements — turned critics into contributors. The project now has 14 contributors and 125+ commits.

14 Contributors on GitHub

Key contributors include @bensig (Ben Sigman, co-author), @milla-jovovich (Milla Jovovich, co-author), @igorls, @adv3nt3, @sheetsync, and @ac-opensource. Community members who caught critical bugs include @panuhorsmalahti, @lhl, and @gizmax (who independently reproduced the 96.6% benchmark on an M2 Ultra in under 5 minutes).