Skip to content

Workspace memory

Your workspace has a memory. When a conversation or an automated agent step produces something worth keeping — a decision, a piece of advice, a commitment someone made, or a fact about your work — it is written down. The next time an assistant in your workspace starts working, it already knows.

Open the Memory tab to see everything your workspace remembers.

Four kinds of entry:

Kind Example
Decision “The launch moves to March 12.”
Advice “Lead with the case study, not the price.”
Commitment “Dana sends the pricing deck to Acme by Friday.”
Fact “Acme’s fiscal year ends in June.”

Entries come from three places:

  • Conversations and workflow steps. After a chat reply or an agent step finishes, the assistant reads what was said and writes down only what a teammate would want recalled later. Small talk, unanswered questions and one-off numbers are left out, and every entry must come from something that was actually said.
  • An assistant asked to remember. An agent with the memory tools attached can record something when you ask it to (“remember that we chose the annual plan”).
  • You. Use Add to memory at the bottom of the Memory tab.

Entries are kept as written, in the language they were said in.

Memory is shared: what is written down is there for every member of the workspace and for their assistants. Two things stay with one person:

  • When the assistant notices that something is about you personally rather than about the work — your own plans, opinions or circumstances — it keeps that entry private to you. Only you see it, only your assistants are given it, and it is not in anyone else’s export. The workspace owner cannot see it either.
  • When you add an entry yourself, tick Only me to keep it private. An assistant you ask to remember something privately does the same.

A private entry shows a Private badge in your Memory tab. To share it later, retract it and add it again without the tick.

When an assistant attached to a chat template, or an agent step in a workflow, starts a run, it is given a short summary of your workspace’s memory: open commitments first, soonest due first, then the most recent decisions, advice and facts. The summary has a fixed size, so the oldest entries drop out of it first — they are still in the Memory tab and can still be found by searching.

Memory belongs to one workspace. An assistant in another workspace never sees it. Within the workspace, an assistant working for you is given your private entries as well; one working for a colleague is not.

A commitment with a due date appears under Open commitments, with its owner and due date; an overdue one is marked. When the due date arrives, the owner gets a “A commitment is due” notification, once. Choose how you receive it in your notification preferences.

When a commitment is delivered or no longer applies, use Mark done or Cancel it. A closed commitment is never reminded about again. The owner, the person who added it, or the workspace owner can close it.

Type into the search box and press Search. Memory is searched by meaning, so “pricing decision” finds “We will price the pro plan at 40 dollars”. Entries that do not relate to what you typed are left out, so a search for something your workspace never recorded finds nothing. If search by meaning is briefly unavailable, the search matches the words instead and says so.

Memory is a record, so nothing is edited in place or deleted outright:

  • Retract removes an entry from the Memory tab, from what assistants are given, and from reminders. It stays in the record, marked as retracted, and is not learned again if the same thing comes up later.
  • To correct an entry, retract it and add the right version.

You can retract entries you added. The workspace owner can retract any entry, including the ones assistants wrote.

Memory is on by default. If your organisation has turned it off, the Memory tab is hidden and nothing is remembered or recalled. A workspace can also be opted out on its own — ask your workspace owner or administrator.

Writing down what is worth remembering, and searching it, uses a small amount of AI processing. It is billed to your workspace like any other assistant activity, as model tokens and embedding tokens — see Usage and billing.