Explore
The whole site, on one page.
Every page, grouped by use, with one sentence to tell whether it is the one you are looking for. And the journal, newest first.
Who it is for
One page per use, written for the person looking for it.
- Students: study a course
The lecturer's PDF, photos of the projector and your Word notes in one memory.
- Teachers: find your courses again
Ten years of courses, exams and papers, asked by notion rather than by year.
- Chat with your documents
A folder, a question, the answer cites the file. Offline if you want.
- A local second brain
Notes, documents, web pages and conversations in vaults, on your computer.
- Import from AFFiNE
Get your AFFiNE documents out as Markdown, images included, in one command.
- Obsidian notes
Stay on Obsidian. Your vault becomes a memory you can ask, no plugin.
- Logseq notes
Stay on Logseq. Your graph becomes a memory you can ask, locally.
- Markdown notes
A folder of .md files, whatever the editor, read where it is.
The product
What it is, how it works, what it costs.
- Manifesto
The thesis: the relationship layer, and why it is an OS.
- How it works
The local architecture, the recall mechanisms, the Neural Map.
- No account, local-first
What leaves your machine, to whom, and how to check without trusting us.
- Download
Windows, macOS, Linux. The complete OS, free, no account.
- License
Free for yourself. 70 $ once for the rest, for life.
- Roadmap
What has shipped, what is next.
- The team
An independent lab, and some friends.
Ecosystem
What plugs into it, and what gets built on it.
- MnemoHub
The app store of Mnemosyne OS.
- The catalogue
The flagship cartridges, real apps built on memory.
- Connect your notes
Obsidian, Logseq, Notion or a Markdown folder: where your notes come from.
- Notion
Keep Notion for building, export, and take your knowledge local.
- Memory for Claude
A persistent, local memory for Claude, over MCP.
- Vertex AI
Your own Vertex AI key, the memory stays with you.
- The memory dashboard
The Neural Map: your memory as a graph you navigate.
- The Engramm
What an engram is, and the license that carries the name.
Developers
Build on memory: SDK, MCP, cartridges.
- Develop on Mnemosyne OS
Where to start: SDK, MCP, cartridges.
- Build a cartridge
The process, vibe-coded or by hand.
- Commission a cartridge
You have the idea, we build the cartridge.
- npm packages
The dev layer is on npm, MIT.
- GitHub ↗
The public repository: open core, SDK and CLI in the clear.
Resources & legal
To read further, and what the law says.
- Explore the site
Every page on one page, with one sentence each.
- Search
One word, and every page on the site that carries it. The index lives in your browser, nothing leaves.
- Documentation ↗
The guide, engine by engine, step by step.
- Journal
What we measured, what we got wrong, what we shipped.
- Benchmark
77.1% under the strict judge on LongMemEval-M, and what it does not claim.
- Lexicon
The words of AI memory, defined.
- Contact
Let's talk.
- Legal notice
Who publishes this site.
- Terms
The terms of sale, fifteen articles.
- Privacy
What this site and the app know about you, which is close to nothing.
The journal
What we measured, what we got wrong, what we shipped.
- A multiplier cannot rescue a zero August 17, 2026
Dense embeddings can't hear a rare proper noun, and doubling a zero is still zero. What worked: a second, local ranking, and the holdout that kept us honest.
- Present is not the same as loadable August 10, 2026
A package can be there, at the right version, and still be unreachable to the code that imports it. How v1.3.7 shipped with no local model.
- We gave personality control of memory. It cost 31 points. August 3, 2026
The most seductive feature we built this year made retrieval worse. The 61-question measurement, the ablation, and why personality shipped as a costume.
- The memory my brain kept asking for July 20, 2026
Why I built a local-first memory OS, not as a product, but as a prosthesis for a brain that ran too fast for the world.
- 72.9% | and the three questions we miss July 15, 2026
We ran LongMemEval on the full-haystack variant, the one nobody shows. The number, the protocol, the levers we refuted, and the three misses we own.