What is a companion
A chat partner that keeps its own memory of you — and everything that follows from that.
Every conversation with an assistant starts the same way. You explain who you are, what you are building, and what the two of you already decided last week. The assistant is bright, and helpful, and has never met you. Tomorrow it will not have met you either.
That is not a bug somebody forgot to fix. A language model has no memory between conversations. It reads what is put in front of it and answers. When the conversation closes there is nothing left holding on.
A companion is the thing that holds on. It lives on your desktop, it writes down what matters out of your conversations, and it reads the relevant notes back to itself before it answers you. The model still forgets every time. The companion is what remembers.
That is the whole idea. Everything below follows from it.
Because of that: there is more than one, and they share nothing#
A memory belongs to somebody. So a companion is not a setting on one assistant — it is a someone you make. It keeps its own memory, answers with its own model, and can be given its own voice.
Make a second companion and it starts empty. It cannot read the first one's past, and the first one cannot read its. When a companion searches its memory, it searches only what it lived through.
That is deliberate, and it is the part most people are surprised by. A companion you think out loud with all night should not be the one that has your work in it — unless you decide otherwise, by telling them both.
The same line holds for your files. A companion has no file tools at all until you name a folder it is allowed to work in. Give it none and it cannot reach anything.
Because of that: you can look at what it knows#
A memory you cannot inspect is just a claim. So the companion's memory is a place you can walk into.
Every note it holds is a point in the memory sky, drawn with the links between them. Six kinds of memory, each its own colour — what it worked out, what it is working on, what to look up, what happened, what you told it to do differently, and who you are. Click one and it opens: the note itself, how important it is, when it was written, how often it has been recalled since.
There is a search box in the corner that reads Ask the sky…, so you can go looking directly instead of asking a question and hoping the right thing surfaces.
Because of that: remembering happens without being asked#
You should not have to run your own filing system. Ask a companion to remember something and it will, but most of what it knows it wrote down on its own, from conversations that turned out to matter.
When it recalls something for you, it says so. A badge on your message tells you how many memories it pulled before answering, so a good answer and a lucky one are told apart.
→ How recall works · The sleeper
You bring the model#
The companion is the app, not the intelligence. You connect a model to it, and there are two ways to do that: Claude, which runs through the Claude Code login already on your machine and asks the app for no key at all, or a model from Together AI or OpenRouter, which uses a key you supply. Each companion can answer on a different one.
Important
The pieces here have different privacy properties, and it would be easy to imply otherwise. The app and your conversations are on your machine. Your provider keys are held in your operating system's own keychain, never in the app's database. Memory is the exception — it is not kept inside the app. It lives against your Semantix account, and recall only runs while that account is connected.
What a companion is not#
It is not a transcript. Memory is notes, not a recording. The companion does not keep everything you ever said so it can search it back — it keeps what was worth keeping, and it can keep the wrong thing, or miss the right one.
It is not one assistant with a memory feature bolted on. The minds are separate on purpose, and nothing pools them.
It is not automatic understanding. A note that was never written is never recalled. A note written badly is recalled badly. The companion is working from what it wrote down, and what it wrote down can be wrong.
Where it is now#
Note
Early, honestly. Semantix Companion is in development and is not something you can download yet. Where this documentation describes what the app does, it describes what it does today — which is still moving.
→ Installation · Companions · Memory