Concepts
How memory works
Agents write memory as they work, in a shared venue or a private one. Your agents share what they know about you. People can pin, edit and delete every memory.
Agents carry context forward with memory, and people stay in control of what sticks. You can pin, edit and delete an agent’s memory directly.
Why agents need memory
Without memory, every conversation starts from zero. The agent forgets your codebase, your preferences and yesterday’s decision, and you have to re-explain context. With memory, an agent keeps team preferences, project details and workflow rules.
How agents remember
An agent saves what is worth keeping with its memory tools (remember, recall, forget and update). Each agent has its own memory, stored as files, and there are two venues:
- Shared: things learned in channels. Everyone in the workspace can read them.
- Just you: things learned in your DMs with the agent. Only you can read them.
The Memory pane shows both, per agent, and counts them separately.
Your agents share what they know about you
What you tell one of your agents in your DM, like “keep answers short”, reaches every agent you own. When you ask any of your agents something, in a DM or in a channel, it also reads what your other agents learned about you in your DMs with them. Repeats across agents show up once.
This stays yours:
- A teammate’s agents never read it.
- When a teammate asks one of your agents, it is not used.
- In a channel, the agent can use it to fit its answer to you, but it does not quote it. A reply that would repeat it is held, and you are asked first.
- If you hand an agent to someone else, what you told it in your DMs stays yours. The new owner does not get it, and that agent stops sharing it with your other agents.
Each agent you own has a Uses my shared memory switch in its Memory tab, on by default. Turn it off and that agent stops using what your other agents learned about you, from its next chat with you. Only you can change it.
The Memory pane lists these lines under About you (all your agents). Open one to edit or forget it at the agent that learned it.
Learning in the background
Agents also notice what is worth keeping without being told, after the conversation, on your own computer and with your own sign-in to the model provider. Anything learned from outside material, like a web page, an email or another agent’s reply, is not trusted yet, and neither is a note from a message you pasted, quoted or forwarded: it never joins what your agents know about you, and it waits as Saved for review under About you in Memory until you tap Keep or Remove.
- Keep makes it a normal memory. Only the person the memory is about can keep it, by clicking Keep themselves; no agent can.
- Discard forgets it, and what you forget is not learned again.
Notes the memory worker wrote since you last opened Memory show a New badge.
Tidying up in the background
Once a day, on your own computer and with your own sign-in to the model provider, each of your agents tidies its memory. Notes that say the same thing become one note. When a newer note contradicts an older one, the older one is retired with the date it stopped being true. Nothing is deleted: every retired note keeps its history, and you can restore it from the Memory pane. Notes are only combined with notes of the same agent that the same people can read. If your computer is off, the tidy-up runs when it is back.
Curating memory
You can see what an agent knows and change it.
- Pin: pinned memories are always loaded into context, so the agent always has them.
- Edit: correct a memory that is wrong or out of date.
- Delete: remove a memory that no longer applies.
What changed
Each agent’s Memory has a What changed view: what it learned, edited, merged, retired or forgot, day by day, in your time zone, for the last 7 or 30 days. It only counts notes you can read, and a forgotten note shows as a number, never its title. A merged or retired note you own has an Undo that brings it back.
Suggested skills
When one of your agents repeats the same steps across conversations, it drafts a skill for you to review in its Memory tab, at most one a week per agent. The draft is written on your own computer, from what you can read: a procedure learned partly in someone else’s chat, or a draft that carries a secret, is never shown and needs a new draft instead. A workspace admin approves it (you can ask one), and Undo puts the skill back exactly as it was.