Memory
The companion writes its own notes, reads them back before it answers, and lets you look at all of them.
A companion remembers the way somebody with a notebook remembers. Not by holding on to everything — by deciding, in the moment, that a thing was worth writing down, and then finding it again later.
The notes are written by the companion, not by you. That single fact explains most of what follows: why memory works without you managing it, why it sometimes keeps the wrong thing, and why there is nowhere in the app to sit down and edit your memories by hand.
Three separate things happen, and it helps to keep them apart.
Recall: before you get an answer#
Every message you send goes past memory on its way to the model. The companion searches its own past for whatever your message is about, and what it finds is handed to the model along with your words. This happens on its own, every time, without being asked.
You can see it happen. While it runs, the line under the conversation reads …is remembering…, and once your message is sent a small 🧠 badge sits under it saying how many memories were pulled — click it and it opens into the actual list, each with its name, its kind, and how strongly it matched.
Note
The badge is there so a good answer and a lucky one can be told apart. If the companion says something that sounds like it knows you, the badge tells you whether it actually looked anything up.
Sometimes a second, quieter search runs as well: when your message is short, the companion also searches against its own last reply rather than against you. Those hits are marked 🪞 in the badge. It is there to catch the companion contradicting itself.
Carving: while you are still talking#
When something durable happens mid-conversation — a decision, a preference, a correction, a fact about you — the companion can write it down there and then. It does this on its own initiative. You can also just ask, in ordinary words: remember that we settled on the second option.
This is deliberately quiet. Memory work never shows up as a tool call in the transcript; the only sign is that presence line. The intent is for it to feel like remembering rather than like filing.
Important
Writing a memory under a name that already exists replaces it. That is how a memory gets updated — and it means a companion can overwrite something it knew, with something it now believes.
The sleeper: after the conversation has grown#
Some things are only visible once a conversation is over. So a companion also reads its conversations back and distils what is left worth keeping.
It happens on its own, on one of two triggers: a dozen new turns, or four new turns and three quiet minutes. Nothing to remember to do. When it finds something, a notification says so — Slept on its own, with what was carved and what was updated. When it finds nothing, it stays silent, because nothing is the normal case.
You can also run it yourself at any point by sending /sleep as a message. That is not a chat
message — it starts the pass immediately, and narrates each stage while it runs.
Looking at all of it#
Everything a companion has ever kept is one screen: the memory sky, a starfield of its memories with the links it drew between them. Six kinds, each its own colour — what it worked out, what you are working on, something to look up exactly, what happened, a correction you gave it, and who you are.
Filter by kind, sink the unimportant ones with the importance slider, or type into Ask the sky… and search the mind directly. Click any memory and it opens: the note itself, how important it is, when it was carved, and how many times it has been recalled since.
What you control#
Memory has a master switch, and turning it off means memory is never consulted at all. Under it,
three smaller ones you can turn off independently: recall cued by your message, the 🪞 self-check, and
the companion's awareness of the current time and of long gaps between messages. The sleeper has its
own switch too — turn it off and distilling is left entirely to /sleep.
What memory is not#
It is not a recording. Memory is notes. It does not keep your conversations so you can read them back.
But raw conversations are searchable, separately. The companion can search your past chats word-for-word, including any history you imported. That is a different thing from memory, and worth knowing you have: distilled notes on one side, the literal transcript on the other.
You cannot edit or delete a memory by hand. There is no editor and no delete button. The only way a memory changes is the companion writing over it. If it has kept something wrong, the fix today is to tell it so and let it carve the correction.
It is not automatic understanding. A note that was never written is never recalled. A note written badly is recalled badly. The companion answers from what it wrote down, and what it wrote down can be wrong.
Where it lives#
Memory is the one part of the companion that is not kept inside the app. It lives against your Semantix account, connected with a token in Settings → Models → Semantix Memory, and recall only runs while that account is connected. The token itself is held in your operating system's own keychain.
→ The memory sky · Recall · The sleeper