How Milla Jovovich Built MemPalace — The Full Story

Act 1: The Frustration

Milla Jovovich is not the person you expect to find on GitHub.

She is, of course, the actress who made Alice an icon across six Resident Evilfilms, 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 stretch of the traditional imagination, a software developer.

But somewhere in late 2025, Jovovich started using AI — ChatGPT, Claude, the usual suspects — not casually, but intensively. She poured thousands of conversations into these systems: business decisions, creative brainstorming, contract reasoning, debugging sessions for projects she was managing. Over months, those conversations became a sprawling archive of her thinking — a digital record of every alternative she considered, every nuance she weighed, every decision she made and why.

Then she'd open a new chat window, and all of it was gone.

Total amnesia. Every session started from zero. The AI that had just helped her reason through a complex negotiation couldn't remember a single word of it the next morning. She was, in her own words, dealing with “a brilliant assistant with permanent short-term memory loss.”

So she went looking for a fix. She tried Mem0, the YC-backed memory layer that has raised $24 million. She tried Zep. She tried several other memory tools in the emerging ecosystem.

Every one of them had the same fundamental problem: 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 throw away the originals. Her reasoning, the alternatives she considered, the nuance that made her decisions actually useful — discarded. Replaced with neat little 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.”

— The philosophy behind MemPalace

That was the moment. Not a technical insight — a human one. The entire AI memory industry was built on the assumption that storage is expensive and context windows are limited, so you need to be smart about what you keep. But Jovovich, coming at 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.

Act 2: The Build

She needed a developer. She found Ben Sigman.

The details of how they connected are sparse — neither has said much publicly beyond their respective social media posts — but the partnership was clear from the start. Jovovich had the vision: an AI memory system that never throws anything away. Sigman had the engineering chops to make it real. And they had a tool that would let them move fast: Claude Code, Anthropic's AI-assisted coding environment.

They spent months building. The design principle was 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 make them searchable.

The architecture borrowed a metaphor from one of the oldest learning techniques in human history: the memory palace. The ancient Greek technique of mentally placing information in rooms of an imagined building, then walking through it to recall what you need. MemPalace made this literal.

The system organized memories into a hierarchy that any human could understand:

Under the hood, it ran on ChromaDB for vector search and SQLite for metadata — both entirely local, both free. No API keys required. No cloud dependency. No subscription. Your memories stayed on your machine.

The name came naturally: MemPalace. And the internet, being the internet, immediately pointed out the missed opportunity: “She should have called it Resident Eval.”

“Multipass.”

— Ben Sigman, launch day tweet

That single-word tweet — a reference to Jovovich's most quotable line from The Fifth Element — became one of the most-shared posts of the launch. Sigman understood the assignment.

Act 3: The Launch

On April 5, 2026, Milla Jovovich pushed MemPalace to GitHub under her own account: github.com/milla-jovovich/mempalace. MIT licensed. Free to use. Free to modify.

What happened next was one of those rare moments where Hollywood and the developer world collide in real time.

Within 48 hours, the repository had over 7,000 stars.

Ben Sigman's launch tweet passed 1.5 million impressions. Threads, Twitter, LinkedIn — every platform lit up simultaneously. Not because another AI memory tool had launched (there are plenty of those), but because Milla Jovovich had a GitHub account, and the code she shipped actually worked.

The reactions were a genre unto themselves:

Brian Roemmele, the tech commentator and founder of The Zero-Human Company, didn't just tweet about it — he deployed MemPalace to 79 employees. A live production deployment, within days of the public launch, at a company that was already betting its operations on AI-first workflows.

The media coverage wasn't limited to tech circles. The novelty of the story — actress ships open-source software, scores higher than VC-funded competitors — crossed over into mainstream tech conversation. It was the kind of narrative that writes itself: the underdog, the unexpected protagonist, the audacious claim.

But the claim was the part that got people thinking.

Act 4: The Controversy

The README made a bold statement: 100% on LongMemEval, the standard benchmark for AI memory systems. A perfect score. The first one ever published. Higher than Mem0, higher than Zep, higher than every well-funded competitor in the space.

Within hours, the scrutiny began.

Penfield Labspublished a critical analysis on Substack with a title that didn't mince words: “None of the benchmark scores are real.” Their detailed examination raised several pointed questions about MemPalace's claimed performance.

The key criticisms, stripped of rhetoric:

• top_k settings: The LoCoMo benchmark test used top_k=50, which critics argued could exceed the actual number of candidates in the test pool — effectively returning everything and trivially achieving a high score.
• Tuning on the test set: Three specific fixes were made to push the score from 99.4% to 100%. The team disclosed this, but critics argued it constituted overfitting to the evaluation set rather than genuine capability improvement.
• README vs. codebase gaps:Some performance claims in the README didn't match what the code actually implemented. The marketing was ahead of the engineering.
• Hybrid vs. raw distinction: The headline 100% required LLM reranking via Haiku — not purely local as the positioning implied. The raw score was 96.6%: still best-in-class, but not 100%.

GitHub issues #27 and #29 became the epicenter of the technical debate. HackerNews threads ran long — a mix of legitimate methodological scrutiny and the inevitable celebrity-meets-tech spectacle. Some commenters carefully dissected the evaluation pipeline. Others couldn't get past the fact that they were reviewing a pull request from the woman who fought zombies in a red dress.

“The benchmarks are debatable. The architecture is interesting. The fact that I'm reading a GitHub issue filed against Milla Jovovich's repository is something I absolutely did not expect to be doing today.”

— A representative HackerNews comment

The controversy wasn't entirely fair, and it wasn't entirely unfair. Aggressive benchmark marketing is endemic to the AI industry — half the papers on arXiv make claims that crumble under independent replication. What made MemPalace different wasn't that its numbers were questioned. It was who was doing the claiming.

For a deeper dive into every benchmark claim and criticism, see our full independent benchmark analysis.

Act 5: What Happens Next

As of this writing, the MemPalace community is growing quickly. The X Community has crossed 161 members. Feature requests are pouring in: integrations with Cursor, GitHub Copilot, Windows support, improved onboarding.

A Chinese developer community has already emerged around the project — GitHub issue #37 is a dedicated thread for Chinese-language discussion and localization efforts. For an open-source project that's barely a week old, that kind of international adoption is notable.

The real question isn't whether MemPalace's benchmarks hold up under perfect scrutiny. The real question is whether the “store everything” philosophy is the right approach to AI memory.

The incumbent tools — Mem0, Zep, and others — all bet on extraction: use AI to pull out important facts, compress the rest, manage a tidy knowledge graph. It's elegant. It's efficient. And it loses information by design.

MemPalace bets the opposite way: keep everything raw, index it intelligently, let search do the work. It's less elegant. It uses more storage. But you never lose a thought you might need later. That's a real design tradeoff, and it's sparked a genuine debate in the AI developer community about what “memory” should mean for machines.

Whether MemPalace becomes the standard or becomes a footnote, it has already accomplished something unusual: it made people rethink an assumption that the entire industry had taken for granted. And it did it because an actress got frustrated enough to build something better.

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