Mnemosyne OS

About Mnemosyne OS

# Mnemosyne OS remembers you. An ordinary assistant treats you as a stranger in every conversation, and forgets everything when the window closes. Mnemosyne OS starts from the opposite idea. **Memory is the operating system, and the assistant runs on top of it.** It lives on your machine. It reads **only the folders and vaults you choose** to open to it. It remembers what you tell it, write and read, six months later and from one model to the next. Over time it learns how you work and adapts to you. The goal is simple: a real relationship with a human, and a companion for each of their projects. ## A different way of answering A word of warning. The way Mnemosyne OS answers you may surprise you. Its harness is the result of months of research. It pushes back on your ideas and your projects, like a partner who challenges you. The way it answers, and the way it is with a human, are its own. We call this **the relationship layer**. One user told us its answers unsettled her, in a good way. We are curious to read about your experience. ## The control room for your agents - Your coding agents show up as cards on your desktop. You see which one is working and which one is waiting, and you reply from its card. - Your Hermes assistant installs in one click. It reads your memory and answers you on Telegram, in text or by voice note. - Before a risky command, the agent asks for your approval on a card. You answer once, for the session, or for good. ## JARVIS mode You drive Mnemosyne OS with your voice and your hands, like JARVIS. - You talk to the orb. It listens, answers out loud and runs your commands: "open my agenda", "full screen". - Through your webcam, you move a window by pinching its corner and zoom with two hands (beta). - Speech recognition, the orb's voice and hand tracking run on your machine. With a local model, everything runs **100% locally**, and every exchange costs **$0 in tokens**. ## Use cases by profile Drop in a PDF, a folder of notes or a codebase. Six weeks later, your question gets an answer grounded in your documents, with its sources. ### Developers Connect your memory to Claude Code, Cursor or any other MCP client. Your agents find your code, your git history, your decisions and your notes. They start with the context instead of starting from zero. ### Students Import a course as a PDF, a Word file or a photo of the lecture screen. The Melete cartridge turns it into a mind map, flashcards and a quiz. Then you ask your course questions in the chat. ### Researchers Every answer cites its sources. A second pass checks each sentence against the documents it found, and flags a sentence with no support. ### Lawyers You cross-check laws against the documents in a case. Switch to 100% local mode, cut the internet and question your documents with the best local models. Your client's file **stays on your machine**. ## Memory cartridges A memory cartridge fetches public knowledge from its official source and stores it in your memory, where you question it from the chat. The first ones cover laws, medicine, science, world history, known vulnerabilities in your code and registered trademarks. Every text keeps its source, its date and its licence. Many more are on the way. ## Technology and sovereignty ### Separate vaults Your memory is organised into vaults, one per domain: code, research, journal, clients. A MAXIMUM vault stays out of ordinary searches, so your journal never ends up in a work answer. ### Search by meaning and by words The engine combines a search by meaning with a search by exact words, and it decides what the model is allowed to see. At night, a background pass links related topics in recent memory. The engine is measured in public: 77.1% (37/48) on LongMemEval-M with the official judge. ### Local first Models, embeddings and search can run entirely on your machine: unplug the cable and keep questioning your documents. Cloud models remain available, with your own keys or in-app credits. Encryption at rest is yours to turn on. No account is needed. ### Several machines, one network P2P connects your Mnemosyne OS installs on your local network, with no server in between, and brings everything together on one desktop. - Each machine's agents appear on your board, through herdr. - You give each computer a role. For example, the PC with the graphics card draws the images the others ask for. - You share a vault read-only, or you keep one together. ### Hardware Mnemosyne OS runs from **8 GB of RAM**, on Windows, macOS and Linux, and adapts its models to your hardware at install time. With **32 GB of RAM** you can comfortably run everything locally, with the largest models. ### For your tools The MCP server runs locally (`npm install -g @mnemosyne_os/mcp`, or one click in Claude Desktop). Your agents search your vaults and follow the same protection rules as you. An agent never reads a vault you did not open to it.

Features

Screenshots & media

No approved screenshots provided.

Reviews

No published reviews yet. Share your experience.

Questions & answers

No published questions yet.