Milla Jovovich has joined Bitcoin entrepreneur and Libre founder Ben Sigman to launch MemPalace, an open-source AI memory system released on GitHub under the MIT license. The repository drew 5,500 stars in three days, putting the project in front of both tech audiences and film fans almost immediately.
Commit history shows direct involvement from Jovovich
According to the source material, the pair spent months building the system with help from Anthropic’s Claude model. Ben Sigman said Claude Opus 4.6 was listed as a code co-author. The Git history shows Jovovich made 4 of the project’s 7 commits through her own GitHub account, including the initial commit tied to core code. She also posted a project video on Instagram to introduce the release.
Perfect score reported on LongMemEval
The project’s technical claims are a major part of the attention it is getting. Information cited from @OneMillion_AI says MemPalace reached 96.6% Recall@5 on the LongMemEval benchmark using local retrieval without external APIs. With optional Haiku reranking enabled, it posted a perfect 100% score across 500 questions. The team described that result as the highest score recorded on the benchmark.
On other tests, MemPalace scored 92.9% on ConvoMem, and the project said that performance was more than double that of the AI memory product Mem0. It also recorded perfect results across all multi-hop reasoning categories in the LoCoMo benchmark.
A four-layer memory architecture and AAAK compression
MemPalace does not follow the standard vector database approach. Instead, it organizes information using a “memory palace” structure. User conversations are arranged into four levels: Wings for people or projects, Rooms for specific topics, Closets for compressed summaries, and Drawers for verbatim dialogue. These sections are linked through Halls and Tunnels for cross-references. The project’s own tests say this design improves retrieval accuracy by 34%.
The repository also introduces AAAK, a lossless compression dialect built for AI agents. The source says it can compress user context from thousands of tokens to about 120 tokens, a ratio of roughly 30x. Because AAAK is structured text, it does not require a special decoder or model fine-tuning, and any large language model that can read text can work with it directly. The system also includes contradiction checks to catch inconsistencies involving names, pronouns, or ages before output.
Code and benchmark scripts are already public
The MemPalace repository has been published together with its benchmark scripts, allowing developers to reproduce and test the claims on their own. Ben Sigman also invited the community to try the project, submit feedback, fork the codebase, contribute changes, and join the official Discord server.