Business context
Tell the assistant about your business once, and save the rules it gets wrong as learnings, so you don't repeat them in every request.
The assistant doesn't know your business: your team names, what you call customers, who approves what, which systems you use. Business context is where you write that down once. The assistant reads it with every request, along with learnings, short rules saved from earlier corrections. Both are shared by everyone in the workspace.
Open it from Settings > Business context, from Business context in the command menu (⌘ K), or from the context ring under the assistant's message box.
Write the context
The Context tab holds one piece of text, up to 6,000 characters. What helps most:
- Teams and roles, and who approves what
- Systems and documents you use
- Your names for customers, suppliers and products
- Limits, deadlines and service levels
Write roles, not people's names or contact details. The text saves when you click away from it, or with ⌘ Enter.
Only an admin can edit the context. Everyone else sees it read-only and can add what they know as a learning instead. If two admins edit at once, the second save is refused with "Couldn't save. Someone else changed it." Reload to see their version, then make your change again.
Save learnings
A learning is one rule of up to 300 characters, such as "Name lanes after teams, never after people." Each one applies to all diagrams, or to one kind such as process diagrams only. Active learnings share a budget of 6,000 characters; turn one off to make room.
There are four ways to add one:
- From the assistant. When you correct something about your business that will hold for other diagrams too, the assistant fixes the diagram and offers the rule as a card under its answer: Remember this? Select Save to keep it, Edit to change the wording first, or Dismiss. Nothing is saved until you choose.
- With
/remember. Type/rememberand the rule in the assistant's message box, for example/remember Refunds over $500 go to the Finance Desk., and press Enter. It's saved straight away and isn't sent to the model. - From a Bad rating. After you write a note under What went wrong?, select Save as learning to turn the note into a rule.
- By hand. Select Add learning on the Learnings tab.
You can edit, turn off or delete your own learnings; an admin can change anyone's. Learnings added by your AI app through the connector show From your AI app, whichever app added them.
What the assistant does with it
The context and the learnings for the diagram's kind go with every request, after the app's own instructions. The assistant treats them as facts about how you work: it uses your names, rules, limits and deadlines instead of generic ones. When your request says otherwise, the request wins, so "for this one, the CFO approves everything" overrides a rule about the Finance Desk.
The context ring's breakdown shows Business context as its own part, so you can see what it costs in tokens. The context is sent as a separate cached part of the request, so the app's instructions stay cached when you change it.
Choose where it goes
Under Send to, an admin picks which providers receive the context and learnings. Claude and ChatGPT are on at first, since neither trains on API requests. Gemini is off: turn it on only if you use a paid Gemini key, because Google may train on free-tier requests. When the selected model's provider isn't allowed, the assistant works without the context and says so once in the conversation.
History, download and deletion
The History tab, for admins, lists every change for a year: who added, edited, turned off or deleted what, and when, with the text before and after. Deleting a learning also wipes its text from the history. When someone is removed from the workspace, an admin can Erase their name from what they added and keep the rules.
The download button in the sheet's header saves the context and learnings as Markdown. Admins can also download them With history as JSON.
The app refuses to save text that looks like a password, API key or private key, and warns about email addresses, phone numbers and card numbers. The context isn't written to the server's logs.
Use it from your AI app
When you connect your AI app, it can read the same context with read_context before it changes a diagram, and save a rule you give it with save_learning. A learning saved that way takes effect at once and shows From your AI app.