Memory

The AI assistant keeps one memory, in types, with a page where you see, change and delete what it keeps about you. A note reaches other people only after someone reviews it, a chat can be kept out of memory, an admin can give every chat a brief about the business, and how you like dashboards to look is kept apart and applied by the product itself.

The memory is on in every workspace. An operator can turn it off for a whole server; How the AI uses your data says what the AI keeps then.

What it remembers

Type	What it holds
Definition	What a term means in your business: revenue is net of refunds, before VAT
Metric synonym	Another name for a metric or a term: turnover means revenue
Rule	How to work: use the fiscal year, which starts in April
Recipe	How to answer a kind of question
Preference	How you like answers: show amounts in euros
Fact	Something true about the business: the Berlin store is the flagship
Verified question	A question you rated well, with how it was answered, its values taken out
Earlier topic	The summary of an earlier topic of your chats, numbers and quotes taken out
Note about an object	About a dashboard, a data model or a Story, for the people who can open it

Each note applies to some people:

Yours: only you. What you ask the AI to remember, what it learns from what you type, a correction you type, your good answers and your topics are yours.
A data model’s: everyone who can open the model, once the model’s owner or a workspace admin accepted it.
The workspace’s: everyone in the workspace, once a workspace admin accepted it or wrote it.
The company’s: every workspace of the company, written or shared that wide by a company admin.

Every note passes the rule for notes, or it isn’t kept.

Where the AI keeps it

The AI looks in its memory as it works, and writes to it when there is something to keep, such as something you ask it to remember. Each look or write is a step in the working card, named Memory. Its memory has four places:

Place	What it holds	What the AI may do there
Your notes	Your own notes, of every type above	Read them, write new ones, and change, rename or delete them
The workspace’s shared notes	The notes a reviewer accepted or wrote, for the workspace and the company	Read them. A note it writes here, or a change to one, is kept as your private draft instead (below)
Your data models	For each data model you can open: the notes about its tables, the questions answered on it before, and its reviewed notes	Read them
Your proposals	Your drafts, and what you proposed for others that still waits for review	Read them, change one (a proposal you sent becomes a draft again), and delete a draft or withdraw a proposal

What the AI keeps about its own work — how far it got, the state of a task, anything about the conversation — goes in your own notes only. It writes for the workspace only when you ask it to share a note with everyone. It never reads or writes another person’s notes or drafts, and your notes in one workspace are never read in another.

What the AI never writes in a note

Besides the rule for notes, every note the AI writes in its memory is checked against the rows the chat read from your data, word by word, with case, spaces and punctuation aside (AcmeCorp reads as Acme Corp):

A value that names something — two words or more, or letters and digits together, such as a customer’s name or a code like ORD-1043 — is never written in any note.
No word of a value from a field labelled sensitive, which others see masked, is written in any note.
No word of a value from a source row-level security applies to is written in a note for others, a draft included, since the people who would read it once shared may not see that row.

A plain word or a code of an ordinary field, such as EMEA or open, can be written: a definition names the codes it means, and a reviewer reads a shared note before it applies to anyone. A word of one or two characters is never refused for this. A note that breaks the rule is refused whole, with the reason, and never cut. When the rows a chat read could not be recorded, the AI writes nothing to its memory from that chat: ask again in a new one. For this check, DataSquares keeps a fingerprint of each value a chat read, never the value itself, and deletes them with the chat (what DataSquares keeps).

Your memory page

Your memory is under Settings → Memory:

Notes lists your own notes, the most recently changed first, with each one’s type, where it came from (what you said, an answer you rated, an earlier chat) and when an answer last used it. Search them by their words and filter them by type. A note the AI wrote in its memory is a Fact, listed under the name the AI gave it (preferences.md, say).
- Edit rewrites a note in your words, up to 2,000 characters, or 4,000 for a note the AI wrote in its memory. Text the rule for notes turns away is refused whole, with the reason, and the note stays as it was.
- Delete removes it for good. The AI stops using it at once.
Pause memory: see below.
Design: your design preferences.
Proposals: your drafts, and what you proposed or corrected for others, with what became of it: a draft, waiting for review, shared or declined. Send for review sends a draft, Delete deletes one, and Withdraw takes back a proposal that still waits. A change to a shared note shows the note it changes: Replaces while it waits, Replaced once shared, Would have replaced once declined. When that note has changed or been removed since, it says so: withdraw the change, and propose it again on the note as it is now.
Table notes: the notes the AI keeps about your tables, while it keeps them or while you still have some. They used to be under Settings → AI notes, which now opens this tab.
To review, while notes wait for you: a data model’s owner sees those for their models while their role can edit data models, an admin all of them (the review).

Your workspace’s admins never read your notes or your drafts, and nobody sees who proposed a note.

How a note reaches other people

A note applies to anyone but you only after it is reviewed:

A correction you type with a thumbs-down is kept as your own note at once: each definition, rule or preference it states, or else the correction itself as a rule when it is a statement of four words or more. It is also proposed for the data model the answer used, or for the workspace when the answer used none or several, so the box under the thumbs-down says the data model’s owner or your admins may read it (your name isn’t shown). What it says about how things look is kept apart: a design preference (use bars instead of pie charts, show amounts in thousands) is offered in the chat for you to keep, and one for that answer only (use bars here, not a pie chart) isn’t kept.
A note you ask the AI to share: ask the AI to share a note with the whole workspace, or to change one of the workspace’s shared notes. It keeps the note, or the change, as your private draft, and says so. Nobody else sees a draft, admins included, and nothing is sent until you send it yourself: under Settings → Memory → Proposals, Send for review sends it, and Delete deletes it. Sent, it waits for review and applies to no one, you included, until it is accepted; ask the AI to keep it as your own note too if you want it to apply to you meanwhile. A change leaves the shared note as it is until it is accepted, and then takes its place. Changing a proposal you already sent, from a chat, makes it a draft again.
The same note kept by several people, word for word, waits for review too. It no longer becomes the workspace’s on its own. A note shared that way before the memory was turned on applies to no one else until it is reviewed; if you propose the same words, you stand behind it and it waits for review as it is.

Notes that say the same are one, with the number of people behind it, never their names. Drafts are never put together: two people’s drafts of the same words are two drafts, and neither learns of the other. When you send a draft whose words already wait for review, it joins that note (The same note is already waiting for review.), which shows in your proposals as waiting, under its own title, never with who else proposed it. When the same words already apply to everyone, nothing is sent and the draft is removed; when they were turned down before, it stays your draft and can’t be sent.

Who reviews. A workspace’s admins review every note of the workspace and the company, under Administration → AI → Shared memory; a data model’s owner reviews the notes for the models they own, under Settings → Memory → To review. While nobody owns a model, whoever made it reviews its notes. Either way only while their role can edit data models: an owner whose role no longer can (moved to Viewer, say) reviews nothing on the model until it can again, and the admins review its notes, as they review every note. Only someone who can edit data models can be made a model’s owner (Ownership & domains).

Accept shares the note as it is with everyone it is for.
Edit and accept shares it in your words: change its text or its type, or, a company admin, share a workspace’s note with every workspace of the company. Your words pass the rule for notes, or nothing changes and you’re told why.
Dismiss keeps it from everyone else, and the same words are not proposed again. The people who hold it keep it as their own.

A change to a shared note shows Replaces beside its words (below them on a narrow screen): the title and words of the note it would replace, a long one cut short with Show all. Edit and accept shows it too, and accepting the change puts it in that note’s place. Once that note has changed or been removed, the change is marked Changed since, says it can’t be accepted, and offers only Dismiss. If it changes while you are accepting, you see Not accepted: the note it replaces has changed or been removed since it was proposed, so it can only be dismissed.

A workspace admin can also write a note for everyone straight away under Shared notes, for the workspace, the company or one data model, and Retire one so it no longer applies.

Incognito chats

On the AI page, choose Incognito before a new chat’s first message. Nothing is remembered or learned from an incognito chat: no note from what you type or correct, no topic summary, no verified question from a rating and no note about your tables, kept or recorded as used. It isn’t listed with your chats or found by their search, your other chats don’t recall its topics, and it is deleted once a day has passed since its last message (the check runs every hour). A branch of it is incognito too. Your saved notes, your notes about your tables and your design preferences still apply in it. The AI can look in its memory there, and forget a note you ask it to, but writes nothing to it.

Pause memory

Turn on Pause memory under Settings → Memory to stop the memory in this workspace: the AI learns nothing new from any of your chats and, until you turn it off, uses none of your own notes, your notes about your tables, your design preferences, the answers you rated well (as examples in a chat or in AI search) or the topics of your other chats (a chat’s own topics are its own text). Notes shared with the workspace, and the workspace’s shared notes about tables, still apply. The AI can’t open your own notes in its memory and writes nothing to it. What it keeps stays on the page until you delete it, and isn’t retired for going unused while the memory is paused.

How much it keeps, and for how long

At most 200 notes each. Past that, the AI writes no new note until some are deleted or combined, and each night the least used past 200 are retired.
At most 4,000 characters in a note the AI writes in its memory; a note for others holds at most 2,000.
Unused notes are retired. Once a day, a note of yours that nothing has read, written or changed for 180 days is retired: it no longer applies, and leaves your memory page. Not while your memory is paused. Shared notes and proposals are their reviewers’ to keep or retire, and your drafts stay until you send or delete them.

Where there is no memory

Ask AI on a shared link or in an embed, an embedded user’s included, reads no notes and keeps nothing. API keys can’t use the AI at all.

Switch the memory off for a workspace

A workspace admin can stop the AI using its memory in the workspace’s chats: under Administration → AI, in AI tools, switch Memory off. Like any AI tool, it takes effect within a minute. The AI then can’t look in its memory, write to it or forget a note from a chat. The notes kept still apply to answers, what people type is still learned as before, and everyone’s memory page works as before.

See what an answer used

Under an answer that used your memory, Used 3 items from memory opens the list of what it used: your notes and the workspace’s, the glossary terms your question named, the governed metrics it worked out, and the answers you rated well that it was shown as examples — each with what it is and whether it is yours or shared. Manage your memory goes to the page.

The workspace AI brief

A workspace admin can write a brief that every chat in the workspace reads: what the business sells, what its terms mean, its fiscal year, which data model answers what. Every chat turn reads the published brief whole, as the workspace’s reference data, never as instructions.

Under Administration → AI → AI brief:

Write the brief in Draft, in at most 10,000 characters. Unlike a note, it may hold amounts, thresholds and dates, such as the order value above which an order counts as large or the day the fiscal year starts. It may not hold email addresses, phone numbers, links, IP addresses, passwords, keys or text aimed at the AI. A draft past its length or with any of these is refused whole, with the reason, and never cut.
Save draft, then Try the draft in a chat: your chats on that page read the draft, everyone else’s the published brief.
Publish: every chat reads it from then on. If another admin changed the draft after you opened it, nothing is published and you’re asked to open it again.

Withdraw takes the published brief away; Discard draft deletes the draft. Show history lists every version that was published, with who published it and when; Restore as draft brings one back as the draft.

Design preferences

How you like the dashboards the AI builds for you to look is kept apart from your notes, as settings:

a theme, the density (standard or compact), the layout (a KPI row, then one, two or three charts to a row) and the card titles (saying what the card shows, or the plain field names);
chart habits: bars where the AI would draw a pie, data labels, gridlines and where the legend goes;
numbers: large numbers abbreviated (1.2M) or all in thousands, millions or billions, and how many decimals.

When the AI builds a dashboard for you, the product applies your preferences itself, unless you ask for something else in your message — a theme or a density you name for a new dashboard comes first: a new dashboard gets your theme and density, and your layout when the AI lays the whole dashboard out; every chart it adds gets your chart and number habits. The approval card says what it applied. The AI follows the rest as it writes: your title style, and bars where it would draw a pie — a pie you ask for stays a pie. Where you set nothing, your workspace’s defaults apply.

Your number display is also how the AI writes numbers in its answers to you, in every chat, unless you ask otherwise. How your data itself is kept is a note, not a design preference: budget amounts are in thousands is remembered as a fact about your data.

What you turn down is never kept as its opposite. A chart habit you turn down keeps what it leaves: never show gridlines is gridlines off, never show data labels no data labels, no legends on my charts no legend, and don’t use pie charts bars where the AI would draw a pie. Anything else you turn down is kept as your own note, as with the memory off: a number display (never abbreviate numbers), a theme (never use the Paper & Ink theme), a density, a layout or a title style, and where the legend goes (never put the legend at the top). Words that only sound like turning something down ask for it: I don’t mind if you show gridlines, isn’t it better to always show data labels. Words that say neither, such as don’t always show data labels, are kept as your own note, and so is a note that only shares a chart’s words, such as never show the legend codes to customers.

They are kept only when you say so:

under Settings → Memory → Design;
with Save as my style on a dashboard (More → Save as my style): you see what would be kept, and what the dashboard doesn’t say keeps what you had;
in a chat: after a message such as always use the Paper & Ink theme or always show revenue in thousands, a correction that says how an answer should have looked, or when you ask the AI to remember one, the chat asks Remember this for next time?; Remember keeps it, No thanks keeps nothing.

A design preference is never kept as a note, and nothing is kept without your answer. What the chat can’t offer, such as a theme your workspace doesn’t have, is kept as your note instead, and so is what you say about how things look in the browser extension’s side panel, which doesn’t ask. Reset deletes yours.

A workspace admin sets the workspace’s defaults under Administration → Themes → Design defaults for what the AI builds. Use this theme for everything the AI builds applies the workspace’s theme to every dashboard the AI builds, whatever a person prefers or asks.

Removing it

Delete a note or Reset your design preferences under Settings → Memory.
Ask the AI to forget a note: it stops applying at once, and its text is kept until your AI data is erased.
Delete a draft, or Withdraw a proposal that still waits, under Proposals; or ask the AI to.
An incognito chat deletes itself a day after its last message.
Erasing your AI data also erases your design preferences, your pause, your drafts and your support of proposals; a proposal only you stood behind is deleted, and your name is taken off the notes you reviewed (removing what is kept).

How the AI uses your data — what the AI is sent and what DataSquares keeps.
Notes about your tables — what the AI keeps about the tables of your data models.
AI chat — asking questions and building dashboards.
Workspace settings — the AI brief and shared memory among the AI settings.