Skip to main content
A new chat does not start from nothing. Before answering, the assistant looks through what you have already discussed and pulls in the few pieces that relate to your message. Three different things get pulled in, and it is worth knowing which is which:
  • Memory — durable facts learned about you. “Prefers TypeScript”, “works in Central European Time”. A handful of short statements, not transcripts.
  • Excerpts — actual passages from earlier chats, quoted back because they resemble what you just asked.
  • Knowledge-base documents — passages from the files you added to a project’s Knowledge base, pulled in when they relate to your message. These are read inside that project’s chats and nowhere else: a knowledge base never reaches a chat outside the project it belongs to.
All three are scoped to you. Nobody else’s chats are ever read, whatever their sharing settings say, and no project’s knowledge base is ever read from outside that project.

​
Seeing what was used

Above the assistant’s reply you will find a line like:
Recalled 3 items
Open it and every item is listed with where it came from. An excerpt carries the chat’s title, the date, and a relevance score; a knowledge-base document carries the filename it was read from, and a score; a memory is shown as the fact itself, with its score. If a reply surprises you, this is the first place to look: it shows exactly what the assistant had in front of it. The line only appears when something was actually recalled. No line means the answer came from the current conversation alone.
Recall costs nothing extra. Finding the excerpts uses one small embedding request per message, which is already part of what you pay for the message.

​
Your memory list

Settings → Profile → Memory lists everything the assistant has learned about you, with the chat each fact came from and whether it applies everywhere or is sealed inside one project. Delete any single fact, or Clear all. Deletion takes effect immediately — the next message will not see it. An empty list usually means exactly what it says: nothing durable has come up yet. Memory is deliberately conservative — it records standing facts, not what you happened to ask about last Tuesday. Short chats, one-off questions and purely technical back-and-forth often produce nothing at all, and that is normal rather than a fault.

​
Keeping a project to itself

Some work should not draw on the rest of your account, and should not leak into it either. Every project carries a memory scope for that, on the project menu in the sidebar: Project only isolates in both directions. A project set that way neither reads from the outside nor contributes to it. A project’s Knowledge base does not depend on this setting. Its documents are always confined to the project they were added to — readable by that project’s chats, invisible to every other chat, on either scope.
Changing the scope does not move memories that already exist, and moving a chat into or out of a project does not move its memories either — each keeps the scope that was in force when it was learned.Chat transcripts behave differently on purpose: they follow the chat. Move a chat into an isolated project and what can be quoted from it moves immediately.

​
Searching your chats

The Search chats box in the sidebar looks at both the words you type and what your chats are about, so a search for “rate limiting” also finds a chat that only ever said “throttling”. Results respect sharing — you only see chats you have access to.

​
Steering recall deliberately

Recall follows resemblance, which is not always what you want. When you know a past thread matters to the conversation you are in — even though nothing you type will resemble it — pin its topic to the chat. See Topics and the graph.