Memory
How the assistant knows what it knows — your documents, the notes it keeps from earlier chats, and the conversation itself. There is nothing to fill in or keep up to date.
What the assistant knows about your organisation is what your documents say, and it reads them when it needs them, the same way a colleague would open the file rather than rely on what they remember. On top of that it keeps short notes from your chats, so it does not ask you the same thing twice. There is nothing to fill in and nothing to keep in step. To change what the assistant knows, change the document, or tell it the new fact.
Where it looks
| Source | What it holds | How it changes |
|---|---|---|
| Your documents | Everything your organisation has written down: policies, procedures, registers, uploaded files with text. The assistant finds them by meaning and reads them in full. | You edit the document, or the assistant edits it when you ask. |
| Its notes from earlier chats | Short notes on facts you told it, what you want, and where things are kept. They carry across chats, so a new chat starts knowing them. | Written when you say something worth keeping. A correction replaces the old note. |
| What a document keeps from chats | A few plain lines on a document: answers someone gave in a chat while that document was being built, that its text does not say yet. Each line says what was said. | Written by the app the moment the answer is given. A new answer to the same question replaces the old one. |
| The conversation | What you said in this chat, and what the assistant's own checks found, so it does not run the same check twice or forget a result. | Ends with the chat. |
When a document's text and one of its kept lines disagree, the text is what counts: the assistant goes by the text and mentions the difference.
It does not ask twice
An answer you give to one of the assistant's questions is kept under that question. If it would ask the same question again, in this chat or a later one, it is stopped and uses your answer instead. To change an answer, just tell it the new one.
Notes that carry across chats
When you tell the assistant something lasting about your organisation, it keeps a note, and a later chat can use it. You do not have to repeat yourself, and you can ask what it has noted.
- It finds notes by meaning and by words. Asking about "who approves policies" can find a note that says "the CEO signs off on our documents", even though the words differ. A note that has not been indexed for meaning is still found by its words.
- A correction replaces the old note. Tell it the earlier fact was wrong and the new note takes over; the assistant uses only the current one. Old notes are never silently edited.
- It does not keep the same note twice. If a note repeats one it already has on the same subject, it says so and keeps the existing one.
- Documents win. If a note disagrees with a document, the assistant tells you and goes by the document. Update the document if the document is the thing that is out of date.
- It says where a fact came from. A note records who said it and when, and the assistant tells you, for example, that you mentioned it earlier this week, instead of presenting it as its own knowledge.
Nothing to set up
A new workspace starts with no interview: after you name it, you go straight to your workspace. The assistant learns about your organisation from the documents you create or upload, and from what you tell it while you work together.