Milla Jovovich Built an AI Memory System That Refuses to Forget

Milla Jovovich is not someone you expect to find on GitHub. She is the actress who made Alice an icon across six Resident Evil films, who played Leeloo in Luc Besson's The Fifth Element, and who has spent three decades as one of Hollywood's most recognizable action stars. She is not, by any conventional measure, a software developer.

And yet, on April 5, 2026, Jovovich pushed code to GitHub under her own account. The repository hit 7,000 stars within 48 hours. It crossed 24,000 stars within a week. Ben Sigman, the developer she partnered with, posted a single-word launch tweet: "Multipass." The internet understood the reference immediately -- and the story went viral in a way that Hollywood and tech rarely collide.

But strip away the celebrity spectacle and something more interesting emerges. MemPalace is a genuine attempt to solve a real problem in AI systems: memory. Not the AI's memory, but yours. And the reason it resonated so loudly isn't just that a famous actress shipped open-source code. It's that she identified a flaw in how the entire AI memory industry thinks -- and built something that challenges the dominant approach.

The Amnesia Problem

Jovovich didn't start using AI casually. She used it intensively -- thousands of conversations with ChatGPT and Claude over months. Business decisions. Creative brainstorming. Contract reasoning. Debugging sessions for projects she was managing. Every conversation was a record of her thinking: alternatives she considered, nuances she weighed, decisions she made and the reasoning behind them.

Then she'd open a new chat window.

All of it, gone. The AI that had just helped her navigate a complex negotiation couldn't recall a single word of it the next morning.

"I was dealing with a brilliant assistant with permanent short-term memory loss." -- Milla Jovovich, Instagram video posted alongside the launch

She tried existing solutions -- Mem0, the YC-backed memory layer that has raised $24 million. Zep. Others in the emerging memory tool ecosystem. They all had the same fundamental flaw: they used AI to decide what was worth keeping.

An LLM would read her conversations, extract what it deemed to be "key facts," compress the rest into summaries, and discard the originals. Her reasoning -- the alternatives considered, the nuance that made her decisions actually useful -- was gone. Replaced with bullet points that missed the point entirely.

"Why should AI decide what I need to remember? It doesn't know what's going to be relevant tomorrow. Nobody does. That's why you keep everything." -- MemPalace philosophy

That is not a technical insight. It is a human one. And it is the core of what makes MemPalace different.

The Architecture: Store Everything, Find It Later

When Jovovich needed a developer, she found Ben Sigman -- coder and CEO of Libre Labs, the Bitcoin lending platform. Sigman had the engineering chops to make her vision real. They spent months building with Claude Code, Anthropic's AI-assisted coding environment, and shipped MemPalace as an MIT-licensed open-source project.

The Memory Palace Hierarchy

The architecture draws from one of the oldest learning techniques in human history: the memory palace. The ancient Greek "method of loci" -- placing information in rooms of an imagined building, then walking through it to recall what you need. MemPalace makes this literal, organizing memories into a hierarchy that any human can navigate:

• Wings -- top-level containers for projects or people
• Rooms -- specific topics within each wing
• Halls -- corridors connecting rooms by memory type (facts, events, discoveries)
• Closets -- 30x compressed summaries using AAAK, a custom lossless format any LLM reads natively without a decoder
• Drawers -- the originals -- verbatim files, the source of truth, never deleted

The Local-First Stack

Under the hood, it runs on ChromaDB for vector search and SQLite for metadata. Both entirely local. Both free. No API keys required. No cloud dependency. No subscription.

This matters for a specific reason. Every major AI memory competitor -- Mem0 at $19--249/month, Zep at $25/month+, Supermemory -- sends your conversations to external services. MemPalace keeps everything on your machine. For a tool whose entire value proposition is preserving private reasoning, that distinction is not incidental.

The core design principle is radical in its simplicity: store everything verbatim. Don't summarize. Don't extract. Don't let an LLM rewrite your memories before filing them away. Just keep the raw conversations, index them properly, and let search do the work.

The Benchmark Controversy

The MemPalace README opens with a bold claim: 100% on LongMemEval, the standard benchmark for AI memory systems. The first perfect score ever published. Higher than Mem0. Higher than Zep. Higher than every well-funded competitor in the space.

Within hours, the scrutiny began.

Penfield Labs published a critical analysis titled "None of the benchmark scores are real," raising pointed questions about the evaluation methodology. The core criticisms:

• The LoCoMo benchmark test used top_k=50, which critics argued could exceed the actual candidate pool size -- effectively returning everything and trivially achieving a high score.
• Three specific fixes were made to push the score from 99.4% to 100%, disclosed by the team but flagged as potential overfitting.
• Some README claims didn't match the actual codebase.
• The headline 100% required LLM reranking via Haiku -- not purely local, despite the positioning.

The raw score -- 96.6% with zero external API calls -- is still the highest published result that requires no cloud services. That is genuinely impressive. But the 100% figure is real, and comes with caveats the marketing does not emphasize.

This is not an isolated problem. Aggressive benchmark marketing is endemic to the AI industry -- half the papers on arXiv make claims that don't survive independent replication. What made the MemPalace controversy different was not that its numbers were questioned. It was who was asking the questions. The spectacle of reviewing a GitHub issue filed against Milla Jovovich's repository was, as one Hacker News commenter put it, "something I absolutely did not expect to be doing today."

Expert Reactions

USC professor of computer science Sean Ren, who also leads Sahara AI, offered a measured perspective: MemPalace represents a different way of structuring how AI systems store information, and it appears to be a general approach that could scale across frameworks. But he cautioned that claims about improved performance have not yet been validated outside controlled tests.

"We need to wait to see how the community reacts when deploying it in real systems." -- Sean Ren, USC professor, to Yahoo Tech

Brian Roemmele, founder of The Zero-Human Company, was less cautious -- he deployed MemPalace to 79 employees within days of launch, at a company already operating on AI-first workflows. Whether that proves the technology or just that the founder was excited by the story is a question only time answers.

What the Industry Got Wrong

The real story beneath the benchmarks is architectural.

Every major AI memory product -- Mem0, Zep, Supermemory -- is built on the same assumption: storage is expensive and context windows are limited, so you need to be smart about what you keep. Use AI to extract key facts. Compress everything else. Manage a tidy knowledge graph. It is elegant. It is efficient. And it loses information by design.

Approach	Optimizes For	Trade-off
Extraction-based (Mem0, Zep)	Efficiency, structured knowledge	Loses information by design
Verbatim storage (MemPalace)	Completeness, recall	Higher storage cost, retrieval complexity

MemPalace bets the opposite way. Keep everything raw. Index it intelligently. Let vector search find what you need. It uses more storage. It is less elegant. But you never lose a thought you might need later.

That is a real trade-off, and it has sparked a genuine debate in the AI developer community about what "memory" should mean for machines. The extraction approach optimizes for a neat system. The verbatim approach optimizes for completeness. Which matters more depends entirely on what you're using the memory for -- and that question has never been seriously asked, because everyone assumed the answer was obvious.

Jovovich, approaching the problem as a user rather than an engineer, saw it differently. The "smart" part was the problem. If you let the machine decide what to forget, you've already lost.

"AI only knows what's already been done. It's the humans running it that actually create something unique and different." -- Milla Jovovich

Whether MemPalace becomes the standard or becomes a footnote, it has already accomplished something unusual: it made the industry rethink an assumption it had taken for granted. And it did it because someone who fights zombies for a living got frustrated enough to build something better.

What's Next

The MemPalace community is growing quickly. Feature requests are pouring in: integrations with Cursor and GitHub Copilot, Windows support, improved onboarding. A Chinese developer community has already emerged around the project, with a dedicated GitHub thread for localization efforts -- notable for a tool barely a week old.

• The X community has crossed 161 members
• GitHub issues are being filed and resolved
• The open questions -- benchmark rigor, real-world deployment at scale, compression accuracy tradeoffs -- remain open

But the project is alive, iterating, and attracting contributors who care about the core idea.

The broader question is whether the "store everything" philosophy gains traction in the AI memory space. Incumbent tools have network effects, funding, and established user bases. But MemPalace is free, local, and MIT-licensed -- which means any developer can take it, extend it, or fork it without permission. In the current AI landscape, where memory systems are increasingly a paid abstraction layer on top of your private conversations, that matters.

The best tools often come from people who are angry about the way things work. Milla Jovovich was angry. And now there's a memory system that doesn't throw away your reasoning.