Remember Each User

Open in CMS

In this step your agent starts remembering useful facts about each person between conversations, so they don't have to repeat themselves. Just as importantly, you decide what it may never remember.

In this tutorial you give Atlas, Kirana Group's operations assistant, two memory fields: the site a person works at (home_site) and their department (department). Kirana's approval answers depend on both. Once Atlas remembers them, a procurement officer at the Surabaya site no longer has to say "Surabaya" every time they ask which cost centre a capex request should use.

In the CMS, open Conversation → Personalization in the left sidebar.

Declare what this agent may remember about the people it talks to. Anything outside that list is never stored. That is what keeps the agent relevant, fast, and explainable.


Before you start


How it works

You define a short list of Memory fields: named slots such as home_site, department, or language. When someone says something that fills one of those slots, the agent records it against their profile. In later conversations it reads those values back and uses them.

Three consequences follow from that design, and they are worth understanding before you configure anything:

  • Nothing is captured that you did not ask for. Atlas declares two fields about where people work. It will not quietly start storing someone's favourite football team, because there is nowhere to put it.
  • Reading costs nothing extra. The fields are a fixed, known list, so recalling them is a single lookup rather than a search. No AI call happens on the way in.
  • You can explain it. Every stored value keeps its origin, the verbatim quote it came from, when it was captured, and when it expires. If someone asks "how do you know that about me?", there is a real answer.

Step 1: Turn on Remember things about each person

At the top of the page is the master switch, Remember things about each person. It is off by default:

Off by default. Nothing is stored until you declare what the agent may remember and the person agrees.

If the organization has Privacy Mode on, this switch has no effect. No memory is captured for any agent in the organization, regardless of what's declared here.

Turn it on. That captures nothing by itself: with no fields declared there is nowhere to put anything. Everything else on the page is configuration for this switch, so the rest appears only once it is on.

If your agent still has the retired personalization mechanism enabled, a warning appears here with a Switch it off button. That older mechanism stored data separately, is being retired, and feeds nothing on this page.

Underneath the switch the page shows three more sections, one below the other:

SectionWhat it holds
Memory fieldsThe declared fields themselves: the part you actually iterate on.
Memory settingsThe prompt budget and the two behavior switches. Set once and rarely revisited.
AI modelThe extraction model.

There are no tabs. One Save button commits the settings and the field list together. Editing a field and adjusting a setting is a single save, not two.

Memory section with Remember things about each person switched on, and the note Off by default. Nothing is stored until you declare what the agent may remember and the person agrees.

Step 2: Declare the memory fields

In the Memory fields section, click Add field. The New memory field drawer opens from the right. Fill it in for the first field, home_site:

SettingValue for AtlasWhat it does
Keyhome_siteStable identity. Lowercase letters, digits and underscores. Changing it creates a different field.
LabelHome siteShown to the person this is remembered about.
PurposeTo answer for the site you work at without asking every time.Required. This exact sentence is what the person is told, and it limits what the data may be used for.
TypeTextA short piece of free text. The other types are listed in Field settings.
Retention (days)180After this, the value is erased. It cannot be unlimited. Default 180, maximum 3650.
When it is includedAlwaysAtlas needs the site in most answers, so it is always available.
Priority100When the token budget is exceeded, the highest numbers are dropped first.
How it may be capturedOnly when stated outrightThe default. Atlas stores the site only when the person says it, never from a guess.
When the value changesKeep the newestSomeone who moves to another site just says so.
Lawful basisLegitimate interestChosen by you as the controller, per field.
SensitivityOrdinary personal dataA work location is not special-category data.

The first six settings are in the main part of the drawer. How it may be captured, When the value changes, Lawful basis, Sensitivity and the three switches are behind Show capture and privacy options: click it to open them. Leave Core field off. Leave Show prominently and The person may correct it on. Click Apply.

The memory field drawer, opened here on Home site: Key home_site, Label Home site, Purpose To answer for the site you work at without asking every time., Type Text, When it is included Always, Priority 100 and Retention (days) 180, with the link Show capture and privacy options and the Cancel and Apply buttons. A new field opens the same drawer, titled New memory field

Click Add field again and declare the second field with the same settings, except:

SettingValue for Atlas
Keydepartment
LabelDepartment
PurposeTo name the right Department Head and cost centre in approval answers.

The Memory fields table now shows both fields. The section counts the declared fields against your organization's limit (2 of 12 declared). The columns are Field (the label and the key), Type, Retention (for example 180 days) and Purpose. Fields carry a core chip when they are core fields, and an orange sensitive chip for special-category data. Click a row to edit the field, or open the menu at the end of the row and choose Edit or Remove.

The table shows no values and no fill counts. When you open an existing field, the drawer shows Stored for 1 of 1 people. That is an aggregate count, never individual values, and the line is missing until someone's value is stored.

Memory fields section showing 2 of 12 declared and the Add field button. The table lists Home site (home_site) and Department (department), both Text, kept 180 days, each with its purpose and a ⋮ menu

Field settings

The drawer also offers these options:

SettingOptions
TypeText, Number, Date, Choice (one of a fixed set), Yes / No, or List (a set of items, each with its own history).
When it is includedAlways · When the conversation mentions a keyword · When a particular tool is available · Only on request. On-demand is never injected: the agent fetches it only when it genuinely needs it.
How it may be capturedOnly when stated outright (the default) · May be inferred from context. Explicit-only means nothing is guessed about the person.
When the value changesKeep the newest · Keep the first · Ask before replacing. A correction the person made themselves is never overwritten by a guess.
Lawful basisConsent · Necessary for a contract · Legitimate interest. Chosen by you as the controller, per field.
SensitivityOrdinary personal data · Special category (health, beliefs, …). Disabled unless your organization has enabled special-category data ("Your organization has not enabled special-category data."); choosing it forces the lawful basis to Consent.

Choice and List fields take a comma-separated Options list. List fields additionally choose which items reach the prompt (most recently confirmed, all of them, or closest to the conversation) and how many. The maximum is 25, and it governs how many items reach the prompt, not how many are stored.

Three switches finish the drawer:

  • Core field: always included, and never trimmed by the token budget.
  • Show prominently: given its own place on the person's memory page rather than sitting in the full list.
  • The person may correct it: they can change the value themselves, and their correction is never overwritten by a guess.

The last two change presentation only. Every field always appears in the person's data access and export, whatever is set above.

Apply stages the change; the page's single Save commits it.

Writing a good field

A field earns its place when it is a fact you would act on and it is stable enough to still be true next time. Good candidates are things a person would be mildly annoyed to repeat: the site they work at, their department, the language they prefer. Poor candidates are anything that changes per conversation, such as the amount of the request they are asking about. That is what conversation context is for.

Keep values short; a slot holds a fact, not a note. Prefer Choice over Text where the answers are genuinely finite, so three spellings of the same site do not accumulate. Kirana could make home_site a Choice field that lists its sites; this tutorial uses Text to keep it short. And set priority deliberately: it is the order in which the agent forgets under pressure.

Changing fields later

Editing the schema of an agent that already has profiles is the easiest way to damage data quietly.

ChangeWhat happens
Removing a fieldIts values stop being used immediately and are erased after a 30-day grace period. Until then the field sits under Scheduled for deletion with a Restore button, and re-declaring the same key brings it back rather than duplicating it.
Changing the typeBest avoided. Treat key and type together as the field's identity, the way you would a database column.
Changing the purpose or sensitivityNot cosmetic. Data gathered for one purpose cannot be redirected to another, so a material change is treated like creating a new field.
Changing retentionShortening applies immediately. Lengthening is not retroactive: values already collected keep the retention the person was told about.
Changing the label or descriptionCosmetic and safe.

Memory fields follow your agent's draft and published versions, so you can add a field on the draft and try it before publishing. The stored data is shared between them, though, which means testing a draft agent writes to real profiles. Test with a test account, not a colleague's.


Step 3: Check Memory settings and AI model

Scroll to the Memory settings and AI model sections. For Atlas the defaults are right, so you only check them.

Memory settings section with Prompt budget (tokens) 400 and the note that the organization allows up to 800, Confirm before storing off and Say where it came from on, and below it the AI model section with Personalization AI model set to Organization default (→ Light)

Prompt budget (tokens)

The limit that applies: the agent trims fields by priority once memory goes over it. Your organization allows up to 800.

Default 400. The note under the field shows how high your organization lets you go (here, up to 800). This, not the field count, is the real constraint, because one List field can grow far larger than a single-value one. Two short Text fields like Atlas's fit easily.

How many fields you may declare is not set here. It is an organization setting. See The limits your organization sets below.

Confirm before storing

The agent asks the person before it writes anything down, instead of capturing it silently.

Off by default. Turning it on makes every capture visible in the moment but slower; leaving it off means the person sees it afterwards, on their own memory page. Which is appropriate depends on how sensitive your declared fields are. Atlas leaves it off. The person still agrees once before anything is stored (see What the person sees).

Say where it came from

The agent says "based on what you told me earlier" when something it remembers shapes a reply.

On by default, and worth leaving on. An agent that silently knows things is unsettling in a way that one which explains itself is not.

Personalization AI model

Picks out the details you declared from what the person says. It runs only when a turn contains something the agent should remember.

That last clause is the cost story: turns that clearly contain nothing relevant are filtered out before any AI call, so most turns cost nothing extra. Leave this on Organization default unless you have a reason not to. The job is small and structured, and does not need your most capable model. Below the dropdown the CMS shows which model the default currently resolves to. This setting sits in its own AI model section.


Step 4: Save

Click Save. The switch, the numbers, and the field list save together. If the backend rejects something, it reports every problem at once, quoting the field: for example "purpose is required on 'home_site'".


Step 5: Let Atlas remember you in the preview

Open the preview while signed in with a test account, start a fresh conversation, and tell Atlas where you work:

I'm based at the Surabaya site and I work in Procurement.

Atlas does not store this straight away. It first asks for permission, along the lines of "May I remember those details?", and tells you what it would keep. Agree, for example with:

Yes, please remember that.

Details you gave before agreeing are not stored. The permission covers what you say from now on. Atlas does not go back and save the sentence you said before you agreed. Say the details again after you agree.

Start a new conversation and say it again:

I'm based at the Surabaya site and I work in Procurement.

Consent is asked once per agent, so Atlas does not ask again. It confirms that you are in Procurement at the Surabaya site and asks what it can help with. Both fields are now stored on your profile.

Atlas preview in a new conversation below the welcome message and example questions. The user writes 'I'm based at the Surabaya site and I work in Procurement.' and Atlas confirms that the user is in Procurement at the Surabaya site and asks 'What operational task can I help with?'

Check the result

  1. Start another new conversation and ask a question where the site matters, without naming it:

    Which cost centre should my next capex request use?
    

    Atlas brings up the Surabaya site on its own and names its cost centre, CC-210. Because a request can be for a different project site, Atlas still checks which site this request is for, as its Interaction instructions tell it to do when the site changes the answer.

    Atlas preview in a new conversation. The user asks 'Which cost centre should my next capex request use?' and Atlas replies: 'Which site is this capex request for? The cost centre depends on the project site (for Surabaya, it's CC-210).' citing the KG-F12 checklist, Cost centres
  2. Ask "what do you remember about me?". Atlas should list your home site and department.

  3. Open the Your memory link in the sidebar to see both values with their purpose, expiry, and the quote they came from. In the CMS, open a field to see Stored for ... people, the count of people who have a value.

  4. Publish the agent when the fields behave the way you want.

If a value is missing, check that you repeated the details after agreeing, that you are signed in, and that the field's Purpose and Type were saved.


The limits your organization sets

Two numbers bounding this page live on the organization rather than the agent, under Organization → Personalization. They are set by an organization Owner or Admin.

LimitWhat it does
Memory fields per agentHow many fields any agent of the organization may declare, up to the platform ceiling of 30. Once the list reaches it, Add field is disabled.
Prompt budget ceilingThe most any agent may spend on remembered data, up to 800. Each agent sets its own budget under it.

The field count is deliberately not an agent setting. Its whole purpose is to bound what a contributor may declare, and a limit the contributor could raise themselves would not be a limit. So the agent holds no number of its own, and the organization's is the only one.

Lowering either afterwards is safe for data already captured. Fields already declared are never erased by a limit lowered later; the agent simply cannot add more until it is back under. A prompt budget above a lowered ceiling is brought down to it.

Cost, incidentally, follows the fields people have actually filled rather than the number declared: twelve declared with three filled costs three.


What the person sees

The people being remembered are not passive here, and most of this is their right rather than a courtesy. Signed-in users of an agent with memory on get a Your memory link in the sidebar. It opens a page that lists everything stored about them: each value with "Kept because: {your purpose}", its expiry, and the quote it came from.

There they can correct a field (where you allowed it), forget one, Download a copy, Erase everything, or Pause remembering, which stops reading and writing without erasing anything.

They can do all of it in plain conversation too, by simply asking: "what do you remember about me?", "forget my home site", "stop remembering things about me". That path matters for people on WhatsApp and other messaging channels, who have no page to visit. It works from the very first turn: someone is entitled to ask what you know and be told "nothing".

Consent is asked once per agent. Rather than interrupting with a dialog, the agent raises it in conversation when one of your declared details comes up, tells the person what it would keep and why, and asks. Nothing is written until they answer, and what they said before agreeing is not saved afterwards. The agent does not press if they decline.

Both the link and the conversational controls appear only for signed-in people; anonymous participants have no profile to keep.


Who on your team can see it

By default nobody sees individual values: contributors see only the aggregate counts, such as Stored for 1 of 1 people. That is deliberate. Tuning a schema, the common task, does not require reading anyone's personal data.

Seeing one person's remembered data takes an explicit grant, is reached from that specific participant rather than browsed as a table, and is recorded every time. Two organization-level switches sit above it, both closed by default and changeable only by an organization Owner:

  • Staff may view what agents remember about individual participants: while off, no contributor can open a participant's data, whatever level they hold.
  • Agents may declare special-category memory fields: while off, the Sensitivity dropdown here is disabled.

Those two carry a higher bar than the limits above, which an Admin may also set. The difference is what they decide: the switches say what personal data the organization is willing to process at all, while the limits only bound how much of it an agent may carry.

See Managing Agent Access for the five levels and who may grant them.


Next steps