How the memory works

mnemo is a memory that AI assistants read from and write to. Instead of every app forgetting you the moment a conversation ends, facts, decisions, and things that happened are written down once and recalled later — by any tool that plugs into the same store.

It is local-first: the real copy of your memory lives in a database on your own machine, and cloud copies are mirrors that stay in step. And it is meaning-based: you don't search it with keywords, you search it by what you mean. The rest of this page explains how that second part actually works.

Every memory is passed through an embedding model — mnemo uses Gemini's multimodal embeddings, with Azure OpenAI as a fallback. The model reads the text and returns a vector: a list of 1,536 numbers that together pin the meaning of that text to a precise point in a very high-dimensional space.

Things that mean similar things land near each other in that space, even when they share no words. "The client wants to pay quarterly" and "invoice Northwind every three months" sit close together. That closeness is the whole trick.

When you recall, your question gets embedded the same way, and pgvector — a Postgres extension for vector math — finds the stored memories whose vectors sit closest to it, ranked by similarity. Documents get the same treatment: long files are split into chunks, each chunk embedded, so a single question can pull back both a remembered fact and the paragraph of a PDF that supports it.

Every memory is tagged with one of five types, borrowed from how cognitive science splits human memory. The type says what kind of thing this is, which lets recall ask for the right flavour — the facts, or the how-to, or the record of what happened.

A memory isn't just its text. Each one carries a small set of numbers that decide how strongly it surfaces over time. New memories start at an importance of 0.5 and drift downward at a decay rate — but every time a memory is recalled and used, it gets reinforced, nudging its importance back up and resetting the clock. Useful memories stay sharp; stale ones quietly sink.

Two independent dials control exposure. Visibility answers whose memory this is. Sensitivity answers how guarded the content is. A recall path can hold both at once — e.g. "only my private memories, and nothing above internal."

Because a memory written now is trusted and re-read later, the moment of writing is the dangerous one. Every write passes through a security scanner first — three checks, and a failure blocks the write outright.

Memories aren't only found by similarity — they can be joined by typed edges that record how one relates to another. That lets the store know a newer fact replaces an older one, or that a decision was made in a particular session.

The authoritative copy of your memory is a Postgres database on your own machine. Hosted copies in the cloud are mirrors — they let other machines and voice agents read the same memory, but they never become the source of truth.