Find what was actually said

Circleback keeps notes, action items, insights, and a transcript for every captured meeting. The job is to find the right meeting, read what it records, and separate what was decided from what was discussed.

1. Filter by who and when before what

SearchMeetings matches keywords against titles and notes, and titles rarely contain the topic. Resolve people with FindProfiles and companies with FindCompanies, then search with profiles or domains and the date range the user gave you. Do not invent a date range to narrow results — if the user said "recently", say which window you used. Tags (ListTags) find a recurring series. Keep the intent argument honest: it is the user's question, and it shapes which excerpts the tool returns.

2. Notes first, transcript when the wording matters

ReadMeetings (up to 50 IDs at once) returns notes, attendees, action items, and insights — enough for "what was decided". Use SearchTranscripts to find where a phrase was said across meetings, and GetTranscriptsForMeetings when the exact words matter: a commitment, a number, a quote the user wants verbatim. Quote the line; never present a paraphrase as a quote.

3. Separate decided from discussed

An "agreed" line is a decision. An action item is a recorded follow-up, not proof that anyone agreed to what it follows from; call it a decision only when the notes or transcript show the agreement. A topic in the notes with no owner and no outcome was discussed, not decided; say so. Two meetings that disagree are a finding to report, not something to reconcile silently — the later one is not automatically right.

4. Cite the meeting

Every claim names the meeting and its date. For anything the user may act on — a promise to a customer, a deadline — also say who said it. When nothing matches, state the filters you searched (people, domain, dates, tags) so an empty answer can be checked. A meeting that was never captured leaves only a calendar event, which SearchCalendarEvents can confirm.

Anti-patterns

• Keyword search first. Most meetings are found by attendee and date; keywords narrow, they do not locate.
• Treating AI insights as the record. Insights are generated; the notes and the transcript are what happened.
• Reading one instance of a recurring meeting. A weekly call decides things across weeks; check the neighbours before saying it never came up.
• Paging blindly. Fetch the next page only when the last one was full and the question is still open.