I Met a Movie Star Mila Jovovich — As a Coder
More evidence of the democratization of software
I didn’t expect to meet Mila Jovovich through a GitHub issue.
But there I was last week, deep-diving into her AI memory framework called MemPalace, when I discovered something remarkable: the “Resident Evil” and “Fifth Element” star had created one of the most talked-about AI memory systems of 2026. And she’d done it using Claude Code, the same AI-assisted development environment I use daily.
More remarkably, when I found critical bugs in her benchmark methodology, she responded directly through her Claude Code workflow, acknowledging the issues and implementing fixes. Not through a PR team or engineering intermediaries — Mila herself, using AI-assisted development to debug complex memory retrieval algorithms at 9 AM on a Thursday.
This isn’t a story about a celebrity coding stunt. It’s about something much more profound: we’ve entered an era where outcomes and features drive development, not the technical limitations of writing code.
The MemPalace Phenomenon
In April 2026, Mila Jovovich and developer Ben Sigman released MemPalace, an open-source AI memory system that immediately went viral. Within 48 hours, it had over 23,000 GitHub stars. The system claimed to achieve the first perfect score on the LongMemEval benchmark, scoring 96.6% raw recall.
The project represents something unprecedented: a free, locally-running memory system that rivals expensive cloud alternatives like Mem0 ($19-249/month) and Zep ($25+/month). It uses the “memory palace” technique — a classical memory method dating back to ancient Greece — implemented through ChromaDB and SQLite, with zero ongoing API costs.
The technical architecture includes basic Claude Code integration (save hooks every 15 messages and before context compression) and 24 tools via the Model Context Protocol (MCP), making it compatible across multiple AI platforms.
The duo had spent months building it using Claude Code’s AI-assisted development environment. As Sigman noted, he provided “the engineering chops” while Jovovich drove the architectural vision.
When Audits Meet AI-Generated Code
That’s when things got interesting.
As someone who works extensively with AI memory systems — I maintain KuzuMemory, a graph-based memory framework — I was naturally curious about MemPalace’s benchmark methodology. The claimed 96.6% recall rate was extraordinary, especially for a system running entirely locally.
So I dove in.
What I found were several methodological issues that fundamentally undermined the headline numbers. The benchmark adapter was discarding assistant turns in conversation history, causing systematic under-recall on certain question types. More critically, the benchmark wasn’t actually testing MemPalace’s core functionality — it was primarily testing ChromaDB’s raw vector search capabilities.
I filed Issue #242 documenting the assistant turn bug, and Issue #214 showing that the 96.6% score was essentially a ChromaDB score, not a MemPalace score.
Mila’s response was immediate and technically sophisticated:
“Hey @bobmatnyc — I’ve taken a look and ran it through CLI. This is a real bug and it’s urgent. You caught that
benchmarks/longmemeval_bench.pyat lines 189-190 builds each session’s indexed document by concatenating onlyuserrole turns... Fix priority: this must land before any public benchmark re-run.“
She didn’t deflect or dismiss. She debugged the issue herself, identified the exact lines of code causing the problem, explained the downstream impact on other benchmarks, and outlined a detailed fix plan including regression tests.
This wasn’t PR speak. This was an AI-assisted developer engaging seriously with technical criticism.
The Democratization Shift
This interaction crystallized something profound about our current moment in software development.
We’re witnessing the emergence of a new class of builders: technically-minded individuals who understand software conceptually but may not have traditional coding backgrounds. AI-assisted development tools like Claude Code, GitHub Copilot, and Cursor have lowered the implementation barrier to the point where vision and domain expertise matter more than syntax mastery.
Mila Jovovich exemplifies this shift perfectly. Without formal technical education (she left school in 7th grade for modeling), she spent months intensively learning AI-assisted development through Claude Code starting in late 2025. She understood the conceptual framework of memory palaces deeply enough to architect a sophisticated system. Her collaboration with Ben Sigman — CEO of Bitcoin lending platform Libre Labs, who provided the engineering expertise while she drove architectural vision — represents a new model of software development where domain knowledge and AI tool fluency can substitute for traditional programming backgrounds.
The fact that a movie star can release a technically competent, widely-adopted memory framework isn’t a commentary on coding getting easier (though it has). It’s about software development becoming more accessible to domain experts and visionaries who previously couldn’t bridge the implementation gap.
What MemPalace Gets Right
Despite the benchmark issues I uncovered, MemPalace demonstrates genuine technical sophistication. The memory palace metaphor isn’t just marketing — it’s a thoughtful architectural choice that makes AI memory systems more intuitive and debuggable.
The system includes elegant features like per-agent memory “wings” that prevent cross-contamination between different AI assistants. The Claude Code integration hooks are well-designed, automatically triggering memory saves at logical conversation boundaries. The MCP implementation is clean and follows established patterns.
Most importantly, the project tackles a real problem: most AI memory systems are either expensive cloud services or complex local installations. MemPalace provides a middle path that’s both free and relatively easy to deploy.
Through my testing and integration experiments, I learned techniques that improved my own KuzuMemory system. The competitive analysis forced me to think more carefully about memory organization patterns and retrieval strategies. This kind of cross-pollination benefits the entire ecosystem.
The Validation Requirement
But the benchmark controversy highlights a crucial point: democratized software development still requires traditional validation methods.
AI-assisted coding tools excel at implementation but can perpetuate subtle conceptual errors throughout a codebase. The MemPalace benchmark issues weren’t obvious bugs — they were methodological problems that required domain expertise to identify.
This creates an interesting dynamic: AI tools enable rapid development by non-traditional developers, but peer review by experienced practitioners becomes even more critical. The community response to MemPalace’s inflated benchmarks wasn’t hostile — it was collaborative debugging at scale.
Mila’s willingness to engage directly with technical criticism and implement fixes demonstrates the right approach. The democratization of software development doesn’t eliminate the need for technical rigor; it distributes that rigor across a broader community.
The Harness Thesis Validated
This story perfectly validates what I call the “harness thesis” — that we’ve entered an era where AI tool ecosystems matter more than underlying model capabilities.
MemPalace succeeded not because Mila wrote perfect code from scratch, but because she effectively orchestrated Claude Code to implement her vision. The system’s value comes from its architectural choices, integration quality, and user experience — not from novel algorithmic breakthroughs.
Similarly, my ability to audit and improve the system came not from superior coding skills, but from having developed complementary expertise with memory systems and benchmark methodology. The collaboration that emerged — distributed across GitHub issues, with contributors from multiple backgrounds — represents the new model of software development.
We’re not just building different software; we’re building software differently.
Meeting Mila Through Code
In the end, I did meet Mila Jovovich — through our AI Agents, lines of Python code, GitHub issues, and technical discussions about memory retrieval algorithms, mediated by our respective Claude Code workflows. Not the meeting I would have predicted, but somehow more meaningful than a typical celebrity encounter.
She embodies a new archetype: the technical visionary who uses AI tools to implement sophisticated ideas without traditional programming backgrounds. Her willingness to engage with criticism and continuously improve the system demonstrates the collaborative spirit that makes this new era of development possible.
The future of software isn’t just about better AI models or more powerful tools. It’s about enabling more people with domain expertise and creative vision to participate in building the systems that shape our digital world.
And sometimes, that means meeting your childhood movie star idol in a GitHub issue thread, debugging memory palace algorithms together.
Bob Matsuoka is CTO of Duetto and writes about AI-powered engineering at HyperDev.
Related reading:
It’s The Harness Stupid — Why AI tool ecosystems matter more than model capabilities
AI Power Ranking — Tool comparisons and benchmarks for AI practitioners
LinkedIn Newsletter — Strategic AI insights for CTOs and engineering leaders
Referenced Links: