Memory stores facts across conversations. Turn it on for an agent, choose whether its memory tools ask before running, and review what has been remembered.

## Enable memory

Memory is off by default. Open the agent's **Memory** settings and turn on **Remember signed-in users**. The agent then keeps notes about each signed-in person and reads them back in that person's later conversations with this agent.

Memory is private: there is no scope to choose. Facts belong to one person and one agent, and nothing is shared across an organization.

| Conversation                          | What the agent remembers                                    |
| ------------------------------------- | ----------------------------------------------------------- |
| **A signed-in person's conversation** | That person's memories for this agent, read and written.     |
| **An anonymous conversation**         | Nothing — the conversation runs without memory.              |
| **A playground test conversation**    | Nothing — the playground never reads or saves memory.        |

Remembering a person requires [identity](/identity): a conversation with no identified owner has no slice to read or write, so it simply runs without memory.

Save and [publish the agent](/revisions) to apply changes to users. Turning memory off stops its use by that agent; it does not delete stored facts.

## Read remembered facts

Relevant memories are retrieved automatically for the user's message. **Memory search** also gives the agent a tool to look up what it remembers about this person. Its **Allow / Ask** setting controls approval for these explicit searches, not for the automatic retrieval.

## How memories are written

Memories are written two ways, and both ride the one switch.

Facts are extracted automatically in the background: once memory is on, the agent distills a remembered conversation into facts every fifth turn. There is no cadence setting — no **Every turn**, no **Turns between extractions**, and no way to leave memory on with extraction off.

**Save memory** is the agent's own tool for recording a fact directly, used when the person asks it to remember something. It mounts with the switch rather than as a separate grant; its **Allow / Ask** setting is the only choice. **Allow** permits the tool call without approval; **Ask** requires approval.

Extraction uses a model call, so a remembered conversation costs a little more than an unremembered one. See [Usage](/usage).

## Inspect and delete memories

Open the organization's **Memory** page to search stored facts and filter by owner, including shared organization memories or a particular person. Open a memory to inspect its scope, dates, and recorded metadata. Members with memory-write permission can delete it there. Deleting a memory does not prevent the same fact being learned again from a future conversation.

This page manages stored facts, not the agent's prompt or published documentation. See [Prompt](/prompt) and [Agent knowledge](/agent-knowledge) for those sources.