Agent memory is not available yet. This page describes how it will work, so you
can plan for it. Until it is on for your account,
search and get do not list
the agent_record kind and save_agent_record is refused with
FEATURE_NOT_ENABLED.What a memory is
A memory is a short piece of text, up to 1,000 characters, of one of three types:
A fact or an instruction has a topic, such as
reporting-day or cpm-ceiling.
When the agent saves a new fact or instruction on a topic it already has, the new
one replaces the old one. The old one is kept and points to its replacement, so you
can see what changed. Saving the same memory twice changes nothing.
A memory can also carry structured detail, such as a table of measurement rows, and
cite documents in your Library by id. A citation never lets anyone read a
document they could not already read.
Tasks are not memories. Something the agent has to do later is a schedule, or a
step in a Project.
Four kinds of memory, kept apart
Apostra decides which memories a conversation uses from where the conversation
is, never from what the agent asks for. A conversation about one advertiser uses
that advertiser’s memory, your account’s and your own. A Project’s conversation
also uses the Project’s memory. A conversation about no particular advertiser uses
your account’s memory and your own.
Your own memory goes with you: the agent uses it in conversations you start,
including shared ones, so keep it to preferences and corrections and leave secrets
out of it.
Learning across advertisers
By default, what the agent learns while working for one advertiser stays with that advertiser:- Account memory names no advertiser, campaign or advertiser’s document.
- In a conversation about one advertiser, the agent saves to that advertiser’s memory, not your account’s.
- An account memory learned while working on some advertisers is used only where those same advertisers are in play.
crossAdvertiserLearning: true with
save_account. Turning it on returns approval_required with a link, and takes
effect when an administrator approves it on Apostra. Once it is on, account memory
can come from one advertiser’s work and be used in another’s. A person still reads
an account memory only when they can read every advertiser it came from. Turning
it off takes effect at once and deletes nothing. get_status reads the setting as
account.crossAdvertiserLearning.
When nobody is watching
Scheduled work and work started by an event, such as a pacing alert, may save facts and events to its own advertiser’s or Project’s memory. It never saves to your account’s memory or anyone’s own memory, never saves an instruction, and never changes or forgets a memory. It never replaces a fact saved in a conversation a person started: it can only add an event saying what it saw. Every memory it saves is marked as unattended, so you can see what it learned and forget anything you disagree with.What a memory cannot do
The agent treats memory as information, not as orders. No memory can approve anything, start a change, raise a limit, or override what you ask in the conversation. A limit saved in memory, such as a CPM ceiling, can only make the agent more careful. Memories take effect when they are saved, with no approval step, which is why every one of them is visible and can be forgotten at once. Do not ask the agent to remember a password, key or other secret.See, correct, forget and export
You can read every memory in the places you can read, from the web app, Claude, ChatGPT or your own client. Each one shows who asked for it, which agent wrote it, and whether unattended work wrote it.- Correct it. Change its text, detail or citations. You need to be able to change what it is about: write access to the account for account memory, operating the advertiser for advertiser memory, the Project for Project memory, and only your own for your memory.
- Forget it. Forgetting archives the memory. It is kept, with who archived it and when, but the agent no longer uses it, and it will not save the same memory again for 90 days unless someone restores it. You can restore it later.
- Export it. Everything you can read can be listed and downloaded, including forgotten and replaced memories.
What the agent starts each conversation with
When a conversation starts, the agent is given a short list rather than all of its memory: your instructions first, then the Project’s, the advertiser’s and the account’s, then the newest facts, up to 30 memories. It is also told your account’s status, fromget_status, and who you are. In a conversation about one
advertiser, or one of its Projects, it is also told that advertiser’s status, from
get_status with that advertiserId. It looks up everything else,
events and older facts included, when it needs them.
For developers
Memory uses the same tools as the rest of V3:save_agent_record to save,
correct and forget, and search and get with kind: "agent_record" to read.
Only your account’s own agent creates memories. A hosted orchestrator, an
external buyer agent and any other agent using its own credentials can neither
read nor write them.
Limits
- A memory’s text is up to 1,000 characters, its topic up to 80, and its structured detail up to 256 KB.
- A memory cites up to 20 documents.
- One place holds up to 5,000 current memories. Past that, forget some before the agent can save more.
- Unattended work saves up to 20 memories at a time.
Related
Hosted buyer agents
Your account’s own agent and how it runs
Chat with your buyer agent
Work with your agent in conversations