Answers from your records, not from a model's memory
The assistant answers by calling read tools against your own data. When there's no evidence, it tells you that instead of inventing something plausible.
It answers from your work, or it says it doesn't know
Ask it something and it reads your own actions, people and captures — then shows you which records it used. If the answer is not in there, it says so rather than producing a confident sentence.
An assistant that guesses is worse than no assistant, because you have to check everything it says anyway.
One thing: the Thursday rota, which has been with her for two days and is now overdue. You raised it in a voice note on Monday morning. A follow-up is drafted and waiting for your approval.
I don't have anything on that. There is a Maple Court quote due Friday, but nothing recorded about what was agreed.
It answers with tools, or it doesn't answer
The assistant has read access to your records through a fixed set of tools. When none of them return evidence, it says so.
What it can look up
Overdue actions, what's due today, what you're waiting on, everything involving one person or one project, recent decisions, open commitments, risks, and a full-text search across your work. Each of those is a real query against your own data.
Ask "what did I promise Sarah?" and it calls the person lookup and the commitments lookup, then answers from what came back. Ask about something you never captured and you get "I don't have a recorded …" — not a plausible invention.
- Read tools answer questions. They are the only source of facts in an answer.
- Write tools need confirmation. The assistant can propose creating or changing an action; you confirm it in the UI.
- There are no external tools. It cannot send, email, book or post. Not restricted — absent.
- Confidence-aware phrasing. When a match is weak it says it found a possible reference it can't confirm, rather than asserting it.
Hallucinated work facts are measured, not hoped for. The evaluation suite that runs against the extraction and assistant behaviour holds a target of zero invented facts about your work, alongside precision and deadline-extraction targets. How the guarantees are enforced →
Ask it, or say nothing and read the brief
The assistant answers from your own records using tools, not from a general impression of what you might have meant. When it has nothing recorded, it says exactly that.
It can read your work and, with a confirmation, write to it. It holds no scope that would let it message anybody.
See it on your own work.
Zephyr is in development for iOS and Android. Join the early access list.