The agent starts every request with a short memory of your workspace and a list of playbooks; both change only through a proposal you approve or an edit in Settings.
Before you start#
- A connected ad account. Memory and playbooks are per workspace, empty until written.
- Editing them needs the Manage agent capability (owners and admins). See Roles and members.
1. What memory is#
Memory is one short text, at most 2,400 characters, that the agent reads at the start of every request: what you sell, who your customers are, the days you never want spend, a rule the agent once got wrong. It is not a log; the agent does not append to it on its own.
The current text is in Settings › Agent › What the agent remembers, with its version and last change.
2. How memory changes#
There are two ways, and both leave a record.
- The agent proposes. When a request teaches it something worth keeping, it writes a Memory update proposal into Needs your call with the text it has now, the text it proposes and a one-line reason. Approve it and the text is replaced; Dismiss it and nothing changes. Only one memory proposal can wait at a time, and only a member with Manage agent can approve or dismiss it. If you edited the memory yourself after the agent proposed, the stale proposal is refused on approval so your edit is kept.
- You edit. Open Settings › Agent, change the text and choose Save. Your edit takes effect on the next request.
The agent cannot write memory directly: its database seat can only propose, the same rule that keeps it from spending.
3. What a playbook is#
A playbook is a repeatable workflow the agent worked out once and should not re-derive: for example, "weekly creative check: list active campaigns, read the last seven days, flag anything whose cost per lead rose by a third". Each playbook has a name, a one-line description, the steps, and up to six tags.
The agent sees only the names and descriptions in every request and opens a playbook's steps when a request matches. A workspace holds up to 40 active playbooks.
4. How playbooks change#
- The agent proposes a playbook the same way it proposes a memory update: a Playbook proposal in Needs your call, with the steps written out. Approve it to add or update it.
- You retire or reactivate a playbook in Settings › Agent › Playbooks. A retired playbook is kept but never shown to the agent.
Each playbook shows how often the agent used it and when; an unused one is a candidate to retire.
5. Let the agent suggest on its own (off by default)#
Settings › Agent has a switch, Let the agent suggest memory and playbooks after a request, off by default. When on, after a long request (six or more steps, or one that used a playbook) the agent may add one Memory update or one Playbook proposal to Needs your call. Nothing is written until you approve, and the review is skipped when less than a fifth of the day's agent budget is left.
6. What memory and playbooks are not#
- Not a way to approve spend: a playbook that says "raise the budget" still produces a proposal that waits for a human.
- Not shared between workspaces: an agency keeps a memory per client.
- Not read by anyone else: members with Manage agent and the agent; nothing leaves your workspace.
If it doesn't work#
- "Memory proposal already pending" — approve or dismiss the waiting one first; there is one slot.
- Save stays disabled over 2,400 characters — that is the cap. Move detail into a playbook, which holds up to 6,000 characters.
- The agent ignores a playbook — check its description: the agent picks by description, so say what the playbook is for, not just its name.
- A retired playbook still appears in a proposal — the agent proposed it again from scratch. Dismiss the proposal; retired playbooks stay retired.