What happens to my transcript?

Sonar never joins a call. There is no bot in your meetings and no recording of its own. About once a week it reads, after the fact and through the vendor's own API, the transcripts your meeting tool already made: Google Meet through the Meet API, Zoom from your cloud recordings, Fireflies, Granola and the others through theirs. From each recording it downloads the transcript file and never the video or the audio.

• Each transcript is reduced to plain lines with their speaker labels and cut at the first 6,000 characters. At most five transcripts are read per source per week, newest first.
• The week's transcripts are joined into one block of text, held in memory for the length of the request.
• The block is sent once to the model with the mining instructions, which ask for two to four story-shaped moments and forbid personal names in any of them.
• The moments come back as a title, an angle, one line of factual detail and the meeting they came from. Those are written to your workspace. The block is not written anywhere, and the request ends.

That is the whole path. There is no transcript table, no archive and no cache. A transcript that was still processing on your meeting tool's side is simply absent this week and picked up next time, because there is nothing on Sonar's side to catch up from.

Three things. The moments: title, angle, one line of detail, the kind (a decision, a win, a lesson, a number, a tension) and the source they came from, with the meeting title as provenance for your eyes only. A receipt of the read on the source's row in Settings, Connected sources: how many meetings were read, the titles of up to five, and how many moments came out, so you can check the work. And the token the vendor issued, stored server-side, so next week's read runs without asking again.

The model call goes to Anthropic's Claude. Under the provider's API terms the text sent is not retained for training and your inputs and outputs are not used to train its models. The privacy policy states this in full, and Security lists what every source reads and keeps.

By default a transcript teaches Sonar nothing about how you speak. If you switch on "Learn voice from your sources" in Settings under Voice, the same transcripts are read a second time each week for one purpose: the lines attributed to you by the speaker labels, and no one else's, are distilled into four to six notes about your spoken register, such as sentence length, favourite openers and how you explain. The notes carry a count as evidence and never a quote longer than six words. Each week's notes replace the last, and the lines themselves go with the pass, exactly as the text does in the main read. The switch is off until you turn it on, and the Settings copy names the sources it would read.

The moments are yours. In Settings under Intelligence, Moments, delete one, a selection, everything from one source, everything older than a date, or all of them, or set a retention rule of 30 days, 90 days or 1 year and a daily sweep removes every moment past it, drafted-from ones included. Turning a source off deletes its token and its receipt at once and leaves the moments until you delete them. Deleting your account deletes everything, immediately.