How Unimatrix works

Every AI chat starts from nothing. You explain your project, your preferences, the decision you made last Tuesday and why — and then the conversation ends and all of it evaporates. Open a new chat, or a different AI, and you start explaining again.

Unimatrix is the place that remembers instead. It sits alongside whichever AI you are using and holds the context, so the AI can look things up rather than ask you. Because the memory lives with you and not inside one company's chat history, it still works when you switch from one AI to another, or from your laptop to your phone.

In one sentence: you save things once, and every AI you connect can read them, in every future conversation, on every device.

You can save a memory by asking your AI to remember it, by highlighting text in the browser extension, or by typing it into the dashboard yourself. However it arrives, it lands in the same place.

Saving is instant — you never wait. The tidying up happens a moment later, in the background, and we call the thing that does it the Librarian.

Think of dropping a book at a library desk. You leave; someone else files it. The Librarian:

• Writes a short summary, so a long memory can be skimmed quickly later.
• Picks out keywords and tags.
• Decides which of your Spaces it belongs in — so a note typed during a work session does not end up buried in a personal project.
• Judges how important it looks, which affects how hard it competes to be surfaced later.
• Pulls out the individual facts inside it — small statements like “this project uses Postgres” — so they can be checked and reused on their own.
• Compares those facts against what you have already told it and flags contradictions, so a decision you reversed last month does not keep resurfacing as if it were current.
• Records how confident it is in all of the above, so weak guesses carry less weight than solid ones.

This is also why memory quality improves on its own. The Librarian revisits your memories on a schedule, not just once at save time.

Memories are kept in two places at once. This is the “dual-layer” architecture, and the reason for it is simple: one layer is built for speed, the other for permanence.

You never choose between the two or manage either of them. It is an internal arrangement that exists so recall feels immediate without putting a ceiling on how much you can keep.

When an AI asks Unimatrix what you know about something, we do not simply match words. We run two different kinds of search and combine them — which is what “hybrid retrieval” means.

• Search by meaning.Ask about “the database we picked” and a memory saying “we went with Postgres” will match, even though it shares no words with your question.
• Search by exact words. Meaning search is weak on rare, specific strings — a ticket number, an unusual surname, an error code. Word search catches those.

Each method produces its own ranked list. We merge the two, then take a closer look at the strongest handful and re-rank them properly. A result that both methods liked rises to the top; a result only one of them found still gets a fair hearing. Any standalone facts the Librarian extracted are included too, so a direct answer can be given without hunting through the memory it came from.

Then we account for time. This is recency decay: newer memories are weighted more heavily than older ones, on a sliding scale rather than a cutoff. By default a memory's recency weight roughly halves every ninety days, and memories in an active Space fade more slowly than ones in an archived corner.

Two things are worth being clear about. Decay only ever affects ranking — a memory that has not been touched in years is still there, still searchable, and still wins if it is genuinely the best match. And decay never reaches zero; it flattens out at a floor. Unimatrix does not delete your memories because they got old.

Finally, within a single working session, memories you have already used get a small nudge upward, so a long conversation stays anchored to the thread you are actually on.

The piece that makes this work everywhere is MCP, the Model Context Protocol. MCP is an open standard for letting an AI application talk to an outside tool. It is best thought of as a plug shape: any AI that has the socket can connect to anything with the plug, with no custom integration written for that specific pairing.

Unimatrix is one of those tools. Once you have connected it, your AI can save and look up memories by itself, as a normal part of answering you. You do not copy anything across, and in most cases you do not have to ask — the AI reaches for memory the way it would reach for a calculator.

Because the connection is a published standard rather than a private arrangement, the same memory is reachable from Claude Desktop, Cursor, Windsurf, and a growing list of others. For AI products that do not support MCP — the web versions of ChatGPT, Claude, and Gemini — the browser extension provides the same thing on the page itself.

This is what makes the cross-device story work. Your phone and your laptop are just two more clients reading the same memory, so a conversation you start on one can be continued on the other.

Your memories are always encrypted before they are written to disk, and always sent over an encrypted connection. But there are two different paths into Unimatrix, and they do not offer identical guarantees — so here is the honest version.

• Memories you write in the dashboard's memory composer are encrypted in your browser, using a key derived from an encryption password only you know. The server receives text it cannot read. If you lose that password, nobody — including us — can recover those memories.
• Memories saved through a connected AI, the browser extension, or the API are encrypted on our servers before being stored. They are protected at rest and in transit, but the Librarian has to read the text in order to summarise, tag, and index it — so this path is not zero-knowledge, and staff with database and key access could in principle read it.

That distinction matters when you are deciding what to save. If something is genuinely sensitive, prefer the dashboard composer with an encryption password set. A good habit either way: avoid putting secrets in Space or Location names, which are stored as plain labels so they can be listed and searched.

Full detail lives on the Security & Infrastructure page and the Privacy Policy. If you would rather no third party held any of it, Unimatrix can be self-hosted on your own hardware.

Terms you will run into across the dashboard, the docs, and your AI's tool calls.

Memory

One saved piece of context — a fact, a decision, a preference, a snippet of a conversation. The smallest thing Unimatrix stores.

Space

A container that groups related memories, usually one per project or area of your life. This is the current name for what older parts of the app and docs call a Palace.

Palace

The original name for a Space. You will still see it in some screens, help articles, and API responses — it means the same thing.

Location

A sub-section inside a Palace, used to group memories by topic ("Architecture decisions", "Meeting notes"). Part of the older Palace structure.

Context

Everything an AI needs to know to be useful right now. Unimatrix’s job is to rebuild your context at the start of every conversation instead of making you re-type it.

The Librarian

The background process that files each memory after you save it — summarising, tagging, categorising, and cross-checking it against what you already told us.

Recall

Looking a memory back up. When an AI asks Unimatrix "what do I know about X?", that is a recall.

Embedding

A numeric fingerprint of a piece of text’s meaning. Two memories about the same subject get similar fingerprints, which is what makes search-by-meaning possible.

Semantic fact

A short subject–relationship–object statement the Librarian pulls out of a memory (for example, "Sam — prefers — TypeScript") so it can be checked and reused on its own.

Importance

A score attached to every memory that affects how strongly it competes to be surfaced. It fades slowly if a memory is never used again.

MCP

Model Context Protocol — an open standard that lets an AI application talk to outside tools. It is the plug that connects Claude, Cursor, and others to your Unimatrix memory.

Session

A single working stretch with an AI. Within a session, Unimatrix leans slightly harder on the memories you have already touched, so answers stay on-topic.