How Mnemosyne works, in one page

Mnemosyne is a memory that lives on your machine. It reads documents you point it at, keeps them, and uses them to answer your questions. Everything else is detail around that loop.

The loop, in four steps

You declare where to look. A vault is a memory store, one per domain of your life, and you give it folders to watch. What is in them enters memory, and what changes afterwards is picked up on its own.

It files what it reads. Every document is split, and each piece gets a vector capturing its meaning. That is what lets it be found later without remembering the exact words.

You ask a question. It searches your vaults two ways at once, by meaning and by literal words, keeps the best passages, and hands them to the model along with your question.

The model answers with those in front of it. It does not guess your memory, it reads it. That is why an answer can name the file it came from.

The model is not the memory

They are two separate things, and you choose the second one. The model doing the writing can run on your machine, at a provider, or on your own server. The memory does not move: it stays in its files, with you.

That means changing model loses you nothing, and a local model gives you an installation where nothing leaves at all.

What you govern

Three settings decide the rest, and they answer three different questions.

A vault's attention says who answers: at 0 it goes quiet, without stopping recording.

A vault's protection says where its content is allowed to go: marked MAXIMUM, it drops out of questions that do not name it and its content does not travel to a remote model.

A question's scope says where to look this time: every vault, or a single one you name.

Where to start

Create a vault, give it a folder to watch, wait for indexing to pass, then ask a question about what that folder contains. If the answer finds nothing, the page on empty answers lists the causes in the order to check them.