Mnemosyne OS

Mnemosyne OS is a personal assistant that never forgets. Every other assistant starts each conversation as a stranger. You re-explain your projects, your clients, your constraints, your preferences — every single time. The context window closes, and everything you said is gone. Mnemosyne OS is built on the opposite premise: memory is not a feature bolted onto a chat window, it is the operating system, and the assistant is what runs on top of it. It lives on your machine. It remembers what you tell it, what you write, what you read, and what you build — and it still remembers it in six months, across sessions, across models, across the assistants you plug into it. ## What it actually does Talk to it and it answers with your context already loaded. Drop in a PDF, a folder of notes, a research corpus, a codebase — it ingests, indexes, and can reason over all of it. Ask a question six weeks later and the answer comes back grounded in what you gave it, with the sources it used. Nothing is re-explained. Nothing is lost between sessions. It reads documents out loud, transcribes your voice, watches folders you choose and keeps their contents current, and runs small apps ("cartridges") on top of your memory — a PDF library, a research assistant, a design studio, a restaurant guide you built yourself. ## Why the technology matters **Vaults, not one big bucket.** Memory is split into isolated stores, one per life domain: code, research, journal, clients, notes. Each vault has its own protection level and its own mixing rules. A vault marked MAXIMUM is not blended into ordinary retrieval. A vault marked isolated is never mixed with any other. Your journal does not leak into a work answer because the boundary is structural, not a prompt asking the model to be careful. **Chronicles, not chat logs.** Every memory is stored as content + a semantic type + a vector embedding. That combination is what lets retrieval be sharp instead of fuzzy: the system knows a decision is a decision, a fact is a fact, and a source is a source — so it can rank by meaning rather than keyword overlap. **Retrieval that changes gear.** A quick factual lookup and a deep synthesis across a year of notes are different problems. The retrieval layer adapts its depth and breadth to the question instead of always dumping the same k results into the prompt. The pipeline lives in the main process, not in the model's reach — retrieval decides what a model is allowed to see, so it is not something a clever prompt can talk its way around. **Consolidation while you sleep.** A background pass reviews recent memory, links related material, and strengthens what turns out to matter — the same reason human memory keeps the signal and drops the noise. You wake up to a memory that is better organized than you left it. **Local first, by architecture.** Embeddings and inference can run entirely on your own hardware — local models via GGUF, local embeddings, local vector search, local storage. Cloud models are available when you want the extra capability, either with your own API keys or with in-app credits, and the routing is yours to set. Encryption at rest is available and armed by you. It runs on ordinary consumer hardware — no server, no account required to think. ## Built for your AI assistants, not just for you Mnemosyne OS ships an MCP server, so the assistants you already use — Claude, your IDE, your agents — can query your memory directly instead of starting cold: npm install -g @mnemosyne_os/mcp The MCP server runs locally and connects to the desktop app. Your memory never has to travel to a third-party server for an agent to use it. Agents get semantic search over your vaults, the ability to write new memories back, and the same vault-protection rules you set for yourself — an agent cannot read a vault you did not grant it. There is also a public SDK and CLI on npm for building on top of the memory layer, a cartridge SDK for shipping small apps that run inside the OS, and an IDE bridge extension so your editor can talk to the same memory your assistant uses. ## Who it is for People whose work is cumulative. Researchers with a decade of papers. Developers with a codebase and its history. Writers, consultants, analysts, founders — anyone who has ever thought "I know I worked this out before" and could not find where. Anyone uncomfortable with the idea that their entire professional context lives in someone else's logs. ## Status Mnemosyne OS is in active beta on Windows and macOS, with the interface available in English, French and Spanish (plus German, Portuguese, Russian and Chinese in beta). It is a real product with a real memory engine underneath, not a wrapper around a chat API. The premise is simple: an assistant that starts every conversation knowing who you are, what you are working on, and what you already decided — because it never forgot.