Memory
Agents remember. Memory is what keeps an agent from re-learning your conventions every conversation, and it's the main reason a long-lived agent gets more useful over time.
Who can do what
| Task | User | Agent owner | Admin |
|---|---|---|---|
| Benefit from the agent's memory | ✅ shared in | ✅ | ✅ |
| See an agent's memories | — as Editor | ✅ | ✅ |
| Edit or delete a memory | — as Editor | ✅ | ✅ |
Note the second row: a Viewer cannot see an agent's memories at all. Memory is configuration, and it's kept with the people who can change the agent.
How memories are created
Mostly by the agent, as it works. When it learns something durable — how you like reports formatted, which system is authoritative for a number, a mistake worth not repeating — it saves it.
You can also just tell it: "remember that our fiscal year starts in April."
Four kinds
| Type | For |
|---|---|
| Fact | Something true about your world — a system, a number, a name |
| Preference | How you like things done |
| Procedure | A repeatable sequence the agent worked out |
| Lesson | Something that went wrong once and shouldn't again |
The distinction matters mainly for recall: an agent can ask for just the procedures, or just the preferences, rather than everything it knows.
Reviewing and editing
Agent → Settings → Memories lists them with their type and content. Edit a memory to correct it, or delete it when it's stale.
Worth doing periodically. Memories are injected into the agent's prompt, so a wrong one is actively harmful — an agent confidently applying a convention you abandoned six months ago is worse than one that asks.
Memory is shared, per agent
One memory store per agent, shared by everyone who uses it. If you share an agent, something it learned in your session can surface in a colleague's. That's usually the point — but it's the reason to think before telling a shared agent something personal.
Memory does not cross between agents automatically. Two agents in the same organization each keep their own.
Agents can, however, search across the organization's memories when they need knowledge another agent has learned. So treat memory as organization-visible rather than agent-private.
Memory vs. history vs. state
Three different things that get confused:
- Memory — durable, curated knowledge, injected into every task.
- Session history — the record of past conversations, searchable on demand.
- Task state — an agent's scratchpad for the task in front of it, so it survives a long job without losing the thread. Not durable.