MemPalace visualization

The story everyone's talking about

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

Milla Jovovich and developer Ben Sigman spent months building MemPalace with Claude Code. It scored 100% on LongMemEval — the first perfect score ever. 7,000+ GitHub stars in 48 hours. But not everyone believes the numbers.

Read the Full Story →Who Is Milla Jovovich? →Are the Benchmarks Real?

40.3K+

GitHub Stars

100%

LongMemEval Score

$0

Cost (MIT Licensed)

100%

Local — No Cloud

Why a Hollywood Actress Started Coding AI Tools

Milla Jovovich coding MemPalace — the actress turned AI developer, with a holographic memory palace in the background

Milla Jovovich — the actress behind Alice in the Resident Evil franchise and Leeloo in The Fifth Element — has been using AI tools daily for months. She accumulated thousands of conversations with ChatGPT and Claude: decisions, creative ideas, business reasoning, debugging sessions. Read the full Milla Jovovich backstory →

Then she noticed something infuriating: every time she started a new session, her AI had amnesia. All those decisions, all that context — gone. She tried existing memory tools like Mem0 and Zep, but they had a fundamental problem: they used AI to decide what was worth remembering. The reasoning she needed most was exactly what got discarded.

So she partnered with developer Ben Sigman and spent months building MemPalace using Claude Code. Their radical idea: don't let AI decide what to forget — store everything, then make it findable.

“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.”

— The motivation behind MemPalace, as described by its creators

Read the complete origin story →

The Benchmark Controversy: Is 100% Real?

MemPalace claims a perfect 100% score on LongMemEval — the standard benchmark for AI memory systems. Within hours of launch, the tech community started asking: is this actually legit?

What's credible

  • 96.6% raw score (zero API) is the highest published result requiring no external services
  • Verbatim storage approach is architecturally sound — you can't lose what you never discard
  • The team openly disclosed that 3 fixes (99.4% → 100%) were tuned on failing questions
  • Held-out test score of 98.4% shows strong generalization

What's been questioned

  • 100% hybrid score uses LLM reranking — not purely local
  • LoCoMo test used top_k=50, which may exceed candidate pool size
  • Some README claims don't match the actual codebase
  • AAAK compression may reduce accuracy from 96.6% to ~84.2%

Our take:The 96.6% raw score is genuinely impressive and beats all competitors that don't require API access. The 100% hybrid score is real but comes with caveats that the marketing doesn't emphasize. This is a real project with real innovations, wrapped in aggressive benchmark marketing that has sparked legitimate pushback.

Read our full independent benchmark analysis →

What People Are Saying

The internet reacted. Here are some highlights.

Ben Sigman

@bensig

1.5M+ impressions

“My friend Milla Jovovich and I spent months building MemPalace with Claude Code. First perfect score on LongMemEval. 5,400 GitHub stars in 24 hours.”

Brian Roemmele

@BrianRoemmele

AI industry KOL

“We tested MemPalace at The Zero-Human Company and deployed it to 79 employees. This is a masterpiece.”

Wayne Sutton

@WayneSutton

Viral reaction

“Milla Jovovich launching an AI memory system was NOT on my 2026 bingo card.”

HackerNews comment

denysvitali

Top HN comment

“Yes, that Milla Jovovich (Resident Evil actress). This was definitely not on my 2026 Bingo Card. Missed opportunity to call it Resident Eval.”

Community reaction

@am_will

Viral tweet

“Milla Jovovich has a GitHub. She's co-developed the highest-scoring AI memory system. What a boss.”

Penfield Labs

Substack

Critical analysis

“None of the benchmark scores are real... the LongMemEval 100% was achieved after targeted fixes on specific failing questions.”

How MemPalace Actually Works

The core idea: instead of letting AI decide what's worth remembering (like Mem0 and Zep do), store everything verbatim and use vector search to find it later. It's inspired by the ancient “memory palace” mnemonic technique.

MemPalace architecture

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 in AAAK lossless format

Drawers

Original verbatim files — the source of truth, never deleted

AAAK

Custom 30x compression dialect any LLM reads natively

SystemLongMemEvalCostLocal
MemPalace (raw)96.6%FreeYes
Supermemory~99%PaidNo
Mem0~85%$19-249/moNo
Zep~85%$25/mo+No

Table shows raw scores (no LLM reranking). Sources: LongMemEval paper. See our benchmark analysis for methodology caveats.

What is MemPalace?MemPalace is a free, open-source AI memory system created by actress Milla Jovovich and developer Ben Sigman. As of 2026, it is the highest-scoring free AI memory framework on the LongMemEval benchmark (96.6% raw, 100% hybrid). Unlike cloud-based alternatives like Mem0 ($19–249/mo) and Zep ($25/mo+), MemPalace runs entirely locally with zero API costs and stores all conversation data verbatim rather than using AI to extract summaries. It is MIT-licensed, supports Claude Code, ChatGPT, and Cursor via MCP, and has earned 19,500+ GitHub stars since launch.

1-minute interactive tool

Which AI Memory Framework Fits You?

7 quick questions. Personalized recommendation with cost estimate. No signup needed.

Question 1 of 70%

Where should your AI memories live?

This determines your deployment model and data sovereignty.

GitHub Pulse

What's happening in the MemPalace repository right now.

⭐

40.3K+

GitHub Stars

Growing fast

🔀

5.1K+

Forks

Active community

👀

265+

Watchers

Following updates

💬

385+

Open Issues

Feature requests + bugs

Hottest Issues Right Now

#59 Import support for Cursor, Copilot, Codex, Windsurffeature#50 Multilingual search support (non-English memories)feature#47 UnicodeEncodeError crash on Windowsbug#37 Chinese community deep analysis & discussioncommunity#29 Benchmark methodology questionsdiscussion

Explore MemPalace

🎬

Milla Jovovich

Actress, coder, AI developer — her full profile & GitHub story

📖

The Full Story

How a Resident Evil actress shipped open-source AI

📊

Benchmark Analysis

Independent look at the 100% score controversy

🛠

Setup Guide

Install and configure MemPalace in 5 minutes

⚖

vs Mem0

Feature-by-feature comparison with pricing

Frequently Asked Questions

Who created MemPalace?+
MemPalace was created by Milla Jovovich (the actress known for the Resident Evil franchise) and developer Ben Sigman. They built it over several months using Claude Code after Jovovich became frustrated that existing AI memory systems would decide what to remember for her, discarding the reasoning and context she actually needed.
Does Milla Jovovich actually code?+
Milla Jovovich co-developed MemPalace using Claude Code, an AI-assisted coding tool. She described the project as a collaboration between her vision for how AI memory should work and Ben Sigman's engineering expertise. The project is MIT licensed and available on her GitHub account at github.com/milla-jovovich/mempalace.
What is MemPalace?+
MemPalace is a free, open-source AI memory system that gives large language models persistent, cross-session memory. As of 2026, it holds the highest score on the LongMemEval benchmark (96.6% raw, 100% hybrid) among free tools. It stores all conversation data verbatim — rather than using AI to extract summaries — and uses vector search to retrieve relevant memories. It runs entirely locally using ChromaDB and SQLite, with zero API costs. MemPalace was created by Milla Jovovich and Ben Sigman and is MIT-licensed.
Is the 100% benchmark score real?+
MemPalace claims 100% on LongMemEval in hybrid mode (with optional Haiku reranking) and 96.6% in raw mode (zero API). However, these scores have been questioned by independent reviewers. The 100% hybrid score was achieved after targeted fixes for specific failing questions, and some methodology choices (like top_k settings) have been debated. The raw 96.6% score is generally considered more credible. See our full benchmark analysis for details.
Is MemPalace free?+
Yes — MemPalace is 100% free and MIT-licensed. It runs entirely on your local machine with zero cloud dependencies. The raw search mode requires no API keys and no cost at all. The optional hybrid mode (which achieves the 100% benchmark score) uses a cloud LLM for reranking at approximately $0.001 per query. There are no paid tiers, subscriptions, or usage limits.
How does MemPalace compare to Mem0?+
MemPalace outperforms Mem0 on benchmarks (96.6–100% vs ~85% on LongMemEval) and costs $0 versus Mem0's $19–249/month. The key architectural difference: MemPalace stores everything verbatim and uses vector search, while Mem0 uses an LLM to extract key facts and discards the rest. MemPalace runs fully locally; Mem0 is cloud-first. However, Mem0 has advantages in enterprise support ($24M YC funding), managed infrastructure, and ecosystem maturity. Choose MemPalace for accuracy and privacy, Mem0 for enterprise needs.
How do I install MemPalace?+
Install with pip install mempalace, then run mempalace init. It requires Python 3.9+. For Claude Code integration, add the MCP server config to your settings. Full tutorial at mempalace.tech/guides/setup.
What is AAAK compression?+
AAAK is a custom lossless compression dialect that MemPalace uses to compress stored memories approximately 30x. Any LLM can read AAAK natively without a decoder. However, independent testing suggests that using AAAK may reduce retrieval accuracy from 96.6% to approximately 84.2%, which is a tradeoff worth considering.

Try MemPalace Yourself

Free, open-source, installs in under a minute. See if the hype is justified.

Setup Guide →GitHub ↗

Weekly newsletter

AI Memory System Weekly

Framework updates, benchmark news, controversy breakdowns. One email/week.

Subscribe Free →

Free. No spam. Unsubscribe anytime.