The layer almost no one is building
The model answers without knowing you.
NVIDIA builds the compute. The big labs build models that reason over a million tokens, write code and analyze images. All of that is intelligence. Almost nobody builds the relationship.
Every conversation with an AI starts from zero. The model does not remember the six weeks you spent on this problem. It does not remember the approach that already failed, or the decision you made on a Tuesday at 2 a.m., or why. It answers politely and correctly, from nothing.
Mnemosyne OS is that missing layer. It makes an AI know you from one session to the next, from one project to another, over time. It needs a model to exist, and it replaces none. It turns a raw capability into something coherent for one human, over the years.
In real use
It connects what you'd forgotten you said.
Four exchanges: a private life, a codebase, a caseload, a research log. Each one starts with an ordinary sentence. The memory that comes back is months old and shares no keyword with that sentence. The meaning connects them. That is the relationship layer at work.
You mention in passing that you’re not around much.
i just wanted to keep you posted, i don’t chat with you very often
I get it. That’s what I’m here for, in between. It reminds me of a note from 4 months ago. Thomas told you he felt set aside, taken up by your work. Presence matters, even in the rushed or distant moments.
↳ Connected by meaning, not a single shared word · 4-month-old memory
Seven months after the decision, you’re about to undo it. The reason for it was never in the code.
why is there a queue in front of the writer? i want to drop it, it slows everything down
An architecture note from 7 months ago answers exactly this. Two processes were writing to the same store at once and corrupting it. The queue is there to serialize the writes, not for throughput. Your own note ends with this: don’t remove it without putting a lock in its place.
↳ The decision came back with its reason, from a doc you never reopened · 7-month-old note
The same legal question, on another matter, a few months apart.
opposing counsel says the confidentiality obligation survives termination, i’ve argued this before, i just can’t place it
Five months ago, on a different matter, you noted a ruling on exactly that point. Your comment on it: the survival clause only holds where the contract names the obligations it covers. You’d flagged that the agreement in front of you then didn’t name them.
↳ Linked by the legal question, not by the client or the file number · 5-month-old note
A preprint you just opened wakes up a run of your own that had gone quiet.
just read the preprint that landed this morning: they drop the reranker and the score doesn’t move
You did a run of that shape nine months ago. You logged that removing the reranker changed nothing beyond noise, then parked it because it looked like a measurement error. Their result comes from a different corpus. Two runs, same conclusion.
↳ Your own negative result, back on the table the day someone else publishes it · 9-month-old run
The name
Why call it an "OS"?
It has no kernel and no drivers. It does the job of an operating system: it manages resources on behalf of processes that should not have to deal with them. Linux does it for programs, with CPU, RAM, disk and network. Mnemosyne OS does it for AI agents, with different resources.
| What an agent needs | What Mnemosyne OS manages on its behalf |
|---|---|
| Memory | Vaults: SQLite + vector stores, partitioned by domain, encrypted at rest (opt-in) |
| Context | Chronicles + semantic retrieval: the agent never rebuilds its past by hand |
| Compute | Routing across model tiers (eco / standard / premium, local / cloud) based on task complexity |
| Hardware | Real GPU/CPU allocation for local voice (CUDA detection, isolated sidecars): a heavy model never blocks the app |
| Input / output | A protocol of signed intents (query · ingest · forget · focus) instead of raw reads and writes |
| Security | FGAC, scope-limited JWT tokens, Zero-Trust IPC validation |
| Persistence | Cross-session continuity: no cold start on every call |
The "OS for LLMs" analogy comes from a peer-reviewed research paper .
The line
You rent the model. You own the memory.
Plug in the best model in the world. It sees only the few sources Mnemosyne OS hands it, never your disk. Switch providers tomorrow, or stop using them all: your memory does not move a single byte. It stays in ordinary database files, on your disk. Turn on AES-256 encryption when you want it. The key is sealed in your machine's keystore, and only you hold the recovery phrase.
That is the asymmetry. The model is an engine you swap every six months. Your memory is ten years of your life. You rent the first and you own the second.
If you plug in nothing, the engine runs entirely on your machine. It works on the plane, in the tunnel, and on the day the API dies or the company folds. If this project disappears, your files are still on your disk and still readable. The file format guarantees it, with no contract involved.
What a real relationship requires
Identity, history, continuity.
Mnemosyne OS builds all three. They are functions of the relationship, not of the model, and they change what it feels like to work with an AI.
I
Identity: the AI knows who it is assisting
The AI knows your domain, your working style, your values and your constraints, and they travel with every session. The Soul Protocol carries them: a persistent personality profile, injected into the agent. It ships today in MnemoForge, the command-line tool.
II
History: the AI knows what happened
The history is captured in chronicles, and you query it by meaning, not only by date. A raw log cannot answer "how many times this year". So the dream state consolidates the history between sessions, while you are away. It adds and never replaces.
III
Continuity: everything is already loaded
Every new conversation starts with that identity and that history already loaded. You never rebuild the context by hand. There is no cold start.