“The Castle” actress Milla Jovovich open sourced the most powerful AI memory system “MemPalace” in history: the first full score in history, 96.6% recall rate, MIT open source

When you have months of conversations with AI, every technical decision, every debugging process, every architectural discussion: once the session is over, it all disappears. You just have to start all over again. This is not a problem for ordinary users, but a pain point experienced by the Hollywood actress herself. That’s right, it’s Milla Jovovich, who plays “Alice” in the “Poker” series of movies. After talking to AI for several months, she found that all the accumulated decision-making and thinking processes were in vain, and decided to build a memory system that was truly her own. She and her friends used Claude Code and spent several months developing MemPalace, which was officially open sourced on GitHub today and won the industry standard benchmark test.First perfect score in history, triggering heated discussions in the AI ​​community. GitHub has accumulated more than 8,800 stars, has branched nearly a thousand times.

Actress Milla Jovovich open sourced the most powerful AI memory system in history “MemPalace”

Origin: The “determined” dilemma of existing AI memory systems

“I don’t want AI to decide what is worth remembering: I want memories that are truly mine.” After using AI to assist her work for a long time, Milla Jovovich discovered that the biggest problem with existing memory systems is that they allow AI to play the role of “memory librarian” and automatically decide what is worth retaining. This resulted in a lot of “why this was done” and “the context at the time” being discarded, leaving only a dry summary like “user prefers Postgres”. Her solution is straightforward:Store everything so it can be found.

生化危机女主角 Milla Jovovich 刚在 GitHub 开源了一个 AI 记忆系统，在行业标准 benchmark 上拿了有史以来第一个满分

没错，就是那个爱丽丝🤯

她用 AI 对话几个月后，积累了大量的决策和思考结果全丢了，她觉得现有的记忆系统让 AI 决定什么值得记，不是她想要的

于是她和朋友用 Claude Code… pic.twitter.com/RnxyzxN3QP

— Ray Wang (@wangray) April 7, 2026

Core innovation: AI version of ancient Greek memory palace

The name MemPalace comes directly from the famous ancient Greek “Memory Palace” (Method of Loci) mnemonic: a speaker recites an entire speech by placing ideas in various rooms of a virtual building. MemPalace applies this concept to AI memory organization.

The system adopts a “palace” three-layer structure:

• Wing: Every project, every person, every theme has its own wings
• Room: There are different types of rooms in each wing, such as “Authentication”, “Accounting”, and “Deployment”
• Closet and Drawer: Compressed digests point to the original archive, ensuring that all content can be accurately traced back

also,Hall Connect related rooms within the same wing,Tunnel Then connect related topics between different projects across the wings. Instead of having to search the entire database, the AI ​​knows which wing to go to for the answer as soon as it opens its mouth. MemPalace officials stated that the palace-like space structure alone can improve search efficiency. 34%。

AAAK: a lossless abbreviation language designed for AI

Another important innovation of MemPalace is its own AAAK(Another AI Knowledge) compression language, which is a lossless abbreviation format specially designed for AI, uses универсальная syntax to compress the conversation content 30 times, achieving “zero information loss”, but the disadvantage is that it is completely unreadable for humans (and there is no need to read it).

The design goals of AAAK are:Not read by humans, but quickly read by AI. Six months of conversations (about 19.5 million tokens in total) can be compressed into only about 170 tokens, allowing the AI ​​to grasp all the key context when it wakes up. And because AAAK is just structured text,Any model that can read text will work:Claude, GPT, Gemini, Llama, Mistral, no decoder, no fine-tuning, completely local running (SQLite + ChromaDB), no cloud API required.

benchmark score: the first perfect score in history

MemPalace achieved amazing results in the industry standard memory system benchmark test:

The test also covers multiple benchmarks such as ConvoMem and LoCoMo, and the performance exceeds that of most paid products.

Comparing with mainstream solutions: six months of AI dialogue generates approximately 19.5 million tokens, which cannot be accommodated if directly pasted into the context window; if LLM automatic summary is used, it will cost approximately $507 per year. MemPalace achieves near-perfect recall at an annual cost of approximately $10.

MIT open source, purely local operation

MemPalace adopts MIT Open Source License, runs entirely locally, no network connection required, no need to upload data. All information is stored on the user’s own machine, ensuring privacy and security.

Installation is extremely simple:

pip install mempalace
mempalace init ~/projects/myapp
mempalace mine ~/chats/ --mode convos

Supports three data mining modes:projects(code and documentation),convos(conversation export), and general(Automatically categorized into decision, preference, milestone, problem and emotional context).

For AI tools that already support the MCP protocol (such as Claude, ChatGPT, and Cursor), you only need to perform connection settings once, and AI will automatically call MemPalace’s 19 MCP tools in the background to actively review the conversation history for the user. Users no longer need to manually enter search commands.

Reaction: 350,000 views, 8,800 stars

MemPalace quickly gained attention after being open sourced. Ray Wang’s shared post on X (Twitter) has received over 354,000 views、2,600 likes、441 retweets, triggering a large number of discussions among AI developers and general users. Many people are amazed that a Hollywood actor can produce results that surpass the memory products of all AI companies. This is probably a strange phenomenon that only occurs in the AI ​​era. As long as you have an idea, you can create world-shaking software without advanced programming skills.

Conclusion

From “The Evil Dead” to open source on GitHub, Milla Jovovich proves one thing: in this era, imagination is the most important ticket. Real innovation often comes from those who best understand “where the pain points are” – not the technology itself.

Of course, whether MemPalace’s actual performance can continue to withstand testing still needs to be verified by the community and time. But the core concept it proposes—allowing humans to regain control of memory instead of leaving it to AI to decide—is undoubtedly a powerful question about the current direction of AI memory products.

Source

Source: KOCPC Chinese