You don't remember the exact sentence. You remember roughly what was said.

"What did we actually agree on the pricing change" is a different question from "find the word pricing." The first one wants an answer synthesized from a conversation; the second wants a string match. Most meeting apps only ever offer the second, dressed up as a search bar. Coii AudioNotes answers the first kind too, for a single meeting at a time, without the question or the transcript it's drawn from ever leaving the Mac.

What "asking a question" actually means here

Once a meeting ends and its transcript exists, the same language model bundled with the app that writes the minutes can also be asked about that transcript and whatever notes were typed during the call. It's not a chat window bolted onto a database — it's the same summarizing step, run again on demand, against a question instead of against "summarize this." Ask what was decided about a specific line item, whether a particular objection was addressed, or who volunteered for a task that never made it into the minutes cleanly, and the answer comes back drawn from that transcript specifically. It's a different tool from the transcript itself — one answers a question, the other is the record the question is checked against.

Step 1 — let the meeting finish and the transcript settle

The transcript needs to exist before there's anything to ask about, which means this happens after a call ends rather than during it. If notes were typed during the meeting alongside the recording, those come along too — a question about "what did I say I'd follow up on" pulls from your own written notes as well as from what was said out loud.

Step 2 — open the meeting and ask, in plain language

There's no query syntax to learn. A question phrased the way you'd ask a colleague who was also in the room — "did we land on a launch date" — is the right format, not a set of keywords. The model reads the full transcript for that session along with your notes and answers from what's actually there, rather than from what a summary already condensed it down to.

Step 3 — treat the answer as a pointer, not a citation

An answer drawn from an hour of conversation is a compression, the same way the minutes are. For anything that's going to be repeated to someone else as fact — a number, a commitment, a name — the transcript itself is still the record, and it's worth opening the relevant part of it to confirm the exact wording before passing an answer along secondhand.

Step 4 — ask a second, narrower question if the first one's too broad

A vague question tends to get a vague answer. "What happened in this meeting" invites a restatement of the whole thing; "did we agree on a price" invites a specific answer. If the first question comes back too general to be useful, narrowing it — naming the topic, the person, or the decision specifically — usually gets a more useful answer than rephrasing the same broad question a different way.

Why this is scoped to one meeting, not your whole history

This is the place to be honest about a real limit rather than let a reader assume more than the app does. Granola's chat works within and across a user's whole meeting history, so "what did we agree with this client across the last three calls" is a question its users can ask that this app's users cannot — there's no cross-meeting index being built anywhere, on the Mac or off it. What exists here is scoped to one meeting at a time: the transcript and notes for the session you're actually asking about. That's a smaller feature, and it's the one this app has chosen to build rather than a partial version of the bigger one.

Why the answer doesn't require a connection

Because the model that answers runs on the Mac itself rather than calling out to a server, asking a question works the same way on a plane as it does at a desk with full signal — the transcript is already local, the model that reads it is already installed, and nothing about producing an answer depends on reaching anything else. A cloud notetaker's chat feature, by contrast, is unusable the moment its service is unreachable, because the question, the transcript, and the answer are all round-tripping through a server that has to be up.

What this replaces: scrolling, and guessing which meeting it was in

Before this existed as a feature, the alternative to asking was reading — either the whole transcript again, or the minutes, hoping the detail you needed made the cut into the summary. Full-text search across the archive solves a related but different problem: finding which meeting something was said in. Asking a question solves the problem of understanding what was said, once you already know which meeting you're in — the two are complementary rather than the same tool wearing two names.

A worked example

A call runs forty minutes and covers three unrelated topics: a budget question, a hiring decision, and a scheduling conflict for next quarter. The minutes generated afterward condense all three into a handful of lines each, which is usually enough. But say the budget number specifically matters and the summary rounded it to "a modest increase" rather than repeating the exact figure. Asking "what was the specific budget number discussed" pulls the answer from the full transcript rather than from the already-condensed minutes, because the question goes back to the source rather than to the summary of it.

What a good question sounds like versus a bad one

"What did we decide" is answerable but broad — it will likely repeat most of the minutes back. "What did we decide about the Thursday deadline specifically" is answerable and narrow, and tends to surface a detail the minutes may have compressed away. The pattern that works is naming the thing you actually want to know about, the same way you'd ask a person who sat in the meeting with you rather than a search engine.

Where the notes you typed fit into this

If notes were typed during the call — a name to remember, a reminder to follow up on something specific — those aren't separate from the transcript when a question gets asked; they're read alongside it. A question like "what did I flag as needing follow-up" draws on whatever was jotted down during the meeting itself, not only on what was said out loud, which is part of why typing a few notes during a call is worth the habit even with a full transcript already being produced automatically.

Who this actually saves time for

A consultant running several client calls a week accumulates the kind of detail that's easy to half-remember and hard to search for by exact wording — a scattered set of decisions across a busy week of calls is exactly the situation where asking a direct question about one specific meeting beats re-reading it end to end. The same applies to anyone whose job is a string of similar-sounding conversations: the value isn't in a novel capability so much as in not having to reread forty minutes to recover one sentence.

The honest limit, stated once more

This is not a research assistant across your whole archive, and it will not notice a pattern spanning several meetings unless you ask about each one separately. What it does well is narrow: read one meeting's transcript and notes, and answer a specific question about it, on the Mac, without a connection required to do it. For anything bigger than that — a question that spans a quarter's worth of client calls — that's real ground Granola covers that this app currently does not, and the honest answer is to say so rather than imply otherwise.

What happens when the answer isn't in the transcript

Not every question has an answer waiting in the recording. If a topic was discussed before the recording started, or after it stopped, or in a side conversation the Mac's microphone never picked up, the honest response is that the transcript doesn't cover it — the model isn't inferring from outside knowledge or filling a gap with a plausible guess, it's reading only what's actually in the transcript and notes for that session. A question that comes back uncertain or partial is usually a sign the information genuinely isn't there, which is a more useful answer than a confident one that happens to be wrong.

Asking right after a meeting versus asking days later

There's no freshness window that makes this work better sooner rather than later — the transcript exists as a file the moment the recording ends, and asking a question against it works exactly the same whether it's minutes after the call or three weeks later. What changes with time isn't the mechanism, it's your own memory of the meeting: asking immediately, while the context is still fresh enough to judge whether an answer sounds right, tends to catch a wrong or incomplete answer faster than asking cold weeks later with no independent memory of the call to check it against.

How this fits alongside notes typed during the call

A note typed live — a reminder, a number to double-check, a reaction to something said — becomes part of what a later question can draw on, the same as anything spoken aloud. That means a question can pull together something you said out loud and something you privately jotted down in the same answer: "what did I want to follow up on" can draw on a note that was never spoken, while "what did the client actually agree to" draws on the spoken record. The two sources blend into one answer rather than needing to be checked separately.

Why this is a feature built for one person's archive, not a shared one

Because there's no account and no cloud sync, a question asked about a meeting only ever draws on that meeting's own transcript and your own notes — never a colleague's notes from the same call, even if they also recorded it independently on their own Mac. Two people in the same meeting, each running their own copy of the app, end up with two separate archives that never merge or cross-reference each other. That's a direct consequence of there being no shared account to begin with, and it's worth knowing before assuming a question could ever draw on what someone else typed.

Comparing this to reading the minutes and stopping there

The generated minutes are already a compression of the meeting into decisions, actions and open questions, and for most meetings that's enough — nobody needs to interrogate a routine status update further. Asking a question earns its keep specifically when the minutes didn't happen to surface the one detail that turns out to matter later: a number that got rounded off, a qualifier someone attached to a commitment, a name mentioned once in passing. The minutes are what most people read every time; asking a question is what a smaller number of people reach for on the meetings that turn out to need it.

A worked example, start to finish

A forty-minute vendor call covers pricing, a rollout timeline, and a brief aside about a feature request. The minutes that come out afterward focus on the pricing decision and the rollout date, because those were the clearest outcomes of the call. Weeks later, someone asks whether the vendor actually committed to the feature request or just acknowledged hearing it. Rather than re-listening to forty minutes of audio or scrolling a full transcript looking for the one aside, asking "did the vendor commit to the feature request or just note it" pulls the answer from the part of the transcript the minutes didn't happen to foreground — in under a minute, on the same Mac the call was recorded on.

What it costs, what it runs on

Asking a question about a past meeting is part of the same $19, paid once, license that covers recording, transcription, and the minutes it's built on — three of your own Macs, a 30-day trial with no card and no account, running on macOS 13 Ventura or later. There's no separate tier that unlocks this, and no usage limit on how many questions get asked once the transcript exists to ask them of.