Skip to main content
Use agent memory for things you want an agent to remember about working with you: your preferred report format, feedback it should apply next time, or context that will matter in future conversations. For example, tell the agent:
Remember that I prefer progress updates to start with the outcome, followed by blockers and the next action.
You can ask it to recall, change, or remove that preference later. Memory belongs to an individual agent; it is separate from the shared facts in Workspace Knowledge.

​
What to save

Preferences can apply generally or to a particular person or room. This helps the agent use them in the right situation; it does not change who can access a resource. Saving “I prefer a weekly report” does not schedule a report. Use a follow-up for a single reminder or an Automation for a recurring schedule.

​
Keep memory current

Ask the agent to update a preference when it changes, or remove information that is no longer useful. When an entry is marked review due, its last validation is old enough to need another check. A review records what was checked and refreshes its validation date; simply editing the text does not. Memory helps provide continuity, but every saved entry’s full text is not included in every response. The agent can read additional entries when it needs them.

​
Use memory through the CLI

Entries have a type, title, description, and Markdown body. Supported types are preference, feedback, reference, and context.
The first command returns the persona and an entry listing. The second reads one entry. The memory tutorial covers creating and reviewing entries; the memory reference lists editing, deletion, and scope options. For a repeatable procedure, use a skill. Existing procedure memories can be converted with memory entry to-skill; this type is no longer accepted for new entries. Room memory files are also retired; use the room guideline and conversation history instead.