Hollywood star Milla Jovovich just built a free AI memory system that beats every big tech company. It costs only 70 cents per year and scores a perfect 100 on the hardest memory test in the world.
Yes, that Milla Jovovich. The actress from Resident Evil.
A few days ago, an open-source AI memory system called MemPalace took over the internet. It is now the strongest AI memory tool on the planet.
nude ai generator
But here is the crazy part. Behind this project is not a team of PhD engineers. It is one woman working from home.
That woman is Milla Jovovich. The same person who fought zombies in Resident Evil. The same face from Miu Miu fashion shows. The same mother of three who spends her nights reading about AI.
clothing remover ai
Together with her husband Ben Sigman, she used Claude to build an open-source AI memory system from scratch.
On the industry standard test called LongMemEval, her system beat 500 of the hardest memory challenges in the world. It got a perfect score.
On GitHub, MemPalace already has 17,900 stars and 2,000 forks.
GitHub link: https://github.com/milla-jovovich/mempalace
Even Elon Musk shared it.
Hollywood Star Builds AI With Claude
The story of MemPalace started by accident.
About 20 days ago, Milla and her husband Ben Sigman tried Claude Code for the first time.
Ben is a music producer. He thought Claude Code was just a fun tool for writing code. But Milla saw something different. She saw a way to turn her ideas into real software.
Then something unexpected happened. While testing a game project, the AI forgot everything. It hit a wall.
Milla realized that AI is powerful but has no memory. It cannot remember what you told it yesterday. It cannot connect ideas from different conversations. It starts fresh every time.
She felt this pain herself. Every time she started a new chat with AI, she had to explain her project, her style, her failed ideas, her goals all over again.
So Milla did what she always does when she sees a problem. She turned it into a project. With Ben’s help, she started building a memory system for AI.
Milla used her love for architecture and logic. Ben helped with the code. After six months of work, they launched MemPalace. The name comes from an old memory technique.
So what exactly is MemPalace?
The Memory Palace That Scored 100
MemPalace is based on an idea that is thousands of years old.
In ancient Greece, speakers used a method called the Memory Palace. They imagined a building in their mind. They placed each piece of information in a different room. When they gave a speech, they walked through the building in their mind and remembered everything.
MemPalace takes this idea and turns it into code. It creates a structured space for data:
Every person, every project, every topic gets its own wing in the palace. Each wing has rooms. Rooms hold different types of data. One room for login systems. One room for database choices. One room for design ideas. Everything has its place.
Between the rooms are halls. Halls store summaries, conflicts, and connections. They show how ideas relate to each other.
Between wings, the system builds tunnels. These tunnels connect related topics automatically.
For example, Kai the backend developer has an auth migration in his wing. Driftwood the project also has an auth migration in its wing. The system sees they are the same topic and builds a tunnel between them. Now when you search from one side, you see both views.
Each room has a closet for summaries. Each closet has drawers for the original full conversations. Nothing gets deleted by hand.
When the AI needs to remember something, it does not search everything. It first finds the right wing. Then the right room. Then the right drawer. The search is small and exact.
On over 22,000 real test questions, this method got 60.9% accuracy. With wings and rooms working together, it jumped to 94.8%. That is a 34 point improvement.
In other words, structure is the secret weapon.
All data lives in ChromaDB on your own computer. No API calls. No cloud bills. It is free.
It costs only 70 cents per year to run.
Here is a comparison of how much information different systems keep.
Most AI memory tools take your full conversation history and squeeze it down. For a typical user with 1,950 tokens of chat history, other tools use big models to make summaries. This costs about 507 dollars and loses important details.
MemPalace only sends 170 tokens of key facts to the AI each time. The rest stays in the palace. The system knows when to add new information and when to skip it.
AAAK: The Secret Code That Shrinks AI Memory
MemPalace has one more trick. It is called AAAK.
AAAK is a special way to write AI memory. It compresses information without losing meaning.
Here is an example. In normal English, this story takes about 1,000 tokens:
Priya and Driftwood team lead, Kai backend dev, 3 years, Soren frontend, Maya DevOps, Leo product, building a SaaS data platform, current sprint moving auth to Clerk, Kai comparing Clerk vs Auth0 on price and features.
With AAAK, the same story fits in about 120 tokens.
That is 8 times less data.
The magic of AAAK is that it keeps structure. It works with any large language model. Claude, GPT, Gemini, they all understand it. No special training needed.
48 Hours From Open Source to Internet Sensation
The story is not over yet.
MemPalace went from open source release to internet sensation in just 48 hours.
It all started with one post about AAAK.
AAAK is part of MemPalace. It is a compression language. The official benchmark shows it achieves 30 times compression.
But there is a catch. When tested with a real tokenizer, the numbers look different. The original English text uses 66 tokens. The AAAK version uses 73 tokens. It actually got bigger, not smaller.
On the LongMemEval test, AAAK mode scored 84.2%. The raw mode scored 96.6%. That is a 12.4 point drop.
On the LoCoMo test, the drop was even bigger. 34 points lower.
When people compared MemPalace to standard vector databases like ChromaDB, they found something interesting. MemPalace with its palace structure, wings, and rooms did much better than flat storage. But the AAAK compression actually hurt performance.
This caused a big debate online.
The project claimed to have a knowledge graph with fact checking. But when people looked at the code, they found a simple Python script. No real knowledge graph. No complex reasoning.
Also, Milla and Ben are new to the open source world. They did not delete old tweets. They did not hide their past. They posted the project with a public letter and uploaded everything.
AAAK uses token display tricks. It is a rough compression method. It does not use a real tokenizer.
The 30 times compression claim came from a writing system, not from the memory system. The 34 point improvement on LoCoMo came from the palace structure, not from AAAK.
When people found these issues, they opened tickets on GitHub. They asked questions.
Milla answered with one sentence. She would rather be honest than perfect.
The open source community reacted in an interesting way. After the drama, more people started supporting the project. The 96.6% raw score is real. The structure is real. The honesty is real.
Elon Musk did not kill MemPalace. He gave it a second life.
What This Means for the Future of AI
Tools like Claude, ChatGPT, and Cursor now support MCP. With one command, your AI can use MemPalace.
Within 19 hours of release, people started using it. AI can now remember. And it is free.
The most surprising thing about this project is not the 100% score. It is not the 30 times compression. It is the person behind it.
A Hollywood actress who fought zombies in Resident Evil used Claude to build something that beats the biggest tech companies in the world. And she made it open source and free.
The key lesson is this. Open source wins. Community wins.
Ben made a joke about it. He said they should rename it. Mempalace becomes Multipass.
If you know The Fifth Element, you know that Leeloo’s famous line is about a multipass.
In the end, the multipass works. Just like in the movie.