Can an AI Notetaker Answer Questions About All Your Past Meetings?

The answer is yes, which is why it stopped being the question

Two years ago, asking a meeting notes tool a question about something said six weeks earlier was a real differentiator. Today it is table stakes. Granola has Chat, Fireflies has Global AskFred, Otter has AI Chat within and across meetings, and Fathom has Ask Fathom. If your shortlist is any of the usual four, every one of them will answer a question about your past meetings. So a buyer who evaluates on whether the feature exists has not narrowed anything down.

The useful question is narrower and almost nobody asks it out loud. When you type a question into one of these boxes, what exactly does it read before it answers? Not in principle. In your account, on your plan, with the settings you actually have.

That turns out to be a much smaller and more variable thing than the phrase across all your meetings suggests. Reading the four vendors' own documentation against this one question, four separate limits decide what gets searched, and they work by completely different mechanisms. One is set by where you happen to be standing in the app. One is a lookback window attached to your plan. One quietly removes the detailed layer from under notes you still have. And one caps how many questions you are permitted to ask per month.

The thing all four have in common is the part that should worry an operator. None of them announce themselves while you are typing. The tool does not tell you that it looked at thirty days, or at twenty five meetings, or at a note whose transcript was deleted last Tuesday. It just answers, in the same confident paragraph it would have produced with everything in front of it.

• Scope: what the tool treats as context, decided by where in the app you asked.
• Lookback: how far back your plan permits the question to reach.
• Depth: whether the transcript still exists under the note, or only the summary.
• Allowance: how many questions per month and per meeting you actually get.

Where you are standing when you ask decides what gets read

Granola's documentation is admirably direct about this. Chat works in different contexts depending on where you use it. From the home screen it queries all of your notes across every meeting. Opened inside a single meeting, it answers about that one meeting. Opened on a folder, it searches that collection. You can also tick specific meetings to use as context. The docs put it plainly: your scope is based on where you're chatting, and you no longer toggle transcripts on or off by hand.

That is a sensible design. It is also a trap for the hurried. The question what did we end up deciding about the pricing change behaves differently depending on which window was in front of you when you typed it. Ask it from inside last Thursday's call and you get Thursday's answer, which may be a half-formed version of a decision that was finalised somewhere else. Nothing about the reply flags that it only consulted one meeting, because from the tool's point of view you asked a question about that meeting.

Fathom makes the same split explicit, to the point of selling the two as separate line items. Its feature matrix lists Ask Fathom: AI within a single call and, as a different row, Account-wide Ask Fathom: AI for all calls, which it marks as a beta feature. Two interfaces that look alike and read very different amounts of your history.

The practical habit is to notice the scope indicator before the question rather than after the answer. If you are evaluating tools, this is also the first thing worth deliberately getting wrong in a trial. Ask the same question from three different places and compare the three answers. A tool that gives you a narrower answer without saying so is telling you something about every answer you will get from it later.

• Ask the same question from the home screen, from inside one meeting, and from a folder, then compare.
• Treat an answer with no stated scope as an answer of unknown scope.
• Check whether account-wide asking is a separate feature or a separate plan, because in at least one tool it is both.

How far back the question reaches is a plan setting

This is the limit that does the most damage, because the number is rarely on the page where you are comparing tools, and the four vendors bound it in four different ways.

Granola states it in the Chat documentation: how much note history chat can use depends on your plan, and on the free tier it is the last thirty days of notes, with full history on the paid workspace tiers. The same page repeats it under troubleshooting, advising you to confirm you're on the right workspace plan for the note history you need if chat is not finding things. A reasonable place to put it, and also the place you look only after the feature has already failed you once.

Otter bounds it by count rather than by time. Its pricing matrix gives the free plan a conversation history of the twenty five most recent, with unlimited history above that. Twenty five is not twenty five days. For somebody on four calls a day it is about a working week, and the free tier is where most evaluations happen.

Fireflies bounds it by volume. Storage is measured in total meeting minutes stored, not in files, and the free plan gets four hundred minutes per user, with eight thousand on Pro and unlimited on the business tiers. The documentation is clear that storage does not reset monthly, it is a total cap. What happens at the ceiling is the part worth reading twice: new meetings will still be captured, but you must delete older meetings to access new ones. The tool keeps working, and the price of that is your own history, pruned by hand.

Fathom uses the word for this axis in its own matrix. Paid plans have unlimited lookback for my calls, while the lookback window for team calls depends on plan tier. So on the same tool, a question about your own history and a question about your colleagues' history can reach back different distances.

Four tools, four mechanisms: days, meeting count, stored minutes, and a tiered lookback window. A buyer comparing them on whether you can ask across your meetings records a yes four times and learns nothing.

• Granola free: the last thirty days of notes for chat, full history on paid workspaces.
• Otter free: the twenty five most recent conversations, unlimited above that.
• Fireflies free: four hundred stored minutes per user as a total cap, and at the cap you delete old meetings to keep recording.
• Fathom: unlimited lookback for your own calls on paid plans, with team call lookback set by tier.

The note can survive while the detail underneath it disappears

The third limit is the one almost nobody checks, and it is the most interesting, because it is created by a setting people turn on for good reasons.

Granola lets you auto-delete transcripts on a schedule you choose, from one day up to a year, and enterprise admins can impose one workspace-wide. The feature exists because keeping a verbatim record of every conversation forever is a genuine liability, and shortening that window is often the right call. The documentation is careful to say that notes are not affected and will remain available.

Then it says what that costs you. Chat quality will be degraded, because Granola Chat uses the transcript to answer questions about your meetings, and without it chat responses for affected meetings will be limited to whatever is in your notes. The same page notes you also lose the ability to regenerate notes for that meeting, and that transcript deletion is permanent and irreversible. For enterprise policies there is no cooldown: existing transcripts older than the retention period are deleted immediately when the policy is activated.

Picture what that produces six months later. Your meeting list looks complete. Every call is there, every summary opens, nothing is visibly missing. But the questions you ask about the older half now get answered from summaries alone, while the questions about the recent half get answered from everything that was said. The archive has two different resolutions and the interface shows you one list.

This is the strongest argument for deciding your retention period deliberately and early rather than inheriting it. A privacy setting and a retrieval feature are reaching for the same data, and whichever one you configure first silently constrains the other.

• Ask whether deleting transcripts affects what the assistant can answer, because in at least one tool the vendor says plainly that it does.
• Find out whether a workspace policy applies immediately or after a grace period.
• Decide the retention window before the archive is a year deep, when the trade is still cheap to change.

The asking itself is metered

Here is a limit that caught us by surprise when reading the matrices side by side. Even where the corpus is wide and deep, the number of questions you are allowed is capped.

Otter publishes the clearest version. Its pricing table has a row for the maximum monthly queries for Otter AI Chat: twenty per user on the free plan, with a further limit of three per conversation, rising to fifty on the next tier and two hundred above that. Twenty questions a month is roughly one a working day. Three per conversation means you get an answer, a follow-up, and one more before that meeting is closed to you.

Granola meters a different dimension, gating which models you may choose. On the free tier chat picks the model automatically with no manual selection, while paid workspaces can select standard or thinking models, where the thinking ones are the ones the documentation recommends for complex processing and longer outputs. So the hard questions, which are the reason you wanted this, are steered toward the better models on the plans that cost money.

This matters for evaluation more than for daily use, and it is worth saying why. The way you find out whether cross-meeting asking is good is by asking it twenty bad questions and watching how it fails. If a free trial gives you twenty questions in total, you cannot run that test. You get one careful attempt, which is exactly the condition under which a demo looks better than the product.

• Check the monthly query cap and the per-meeting cap before you plan a trial around this feature.
• Budget your trial questions, or buy a month of the real tier rather than evaluating on a metered free plan.
• Ask which model answers on which plan, because the hard questions are often routed to the better model only above a certain tier.

The corpus stops where your meetings stop

The last limit is the broadest, and here Granola deserves credit for writing down the sentence that most vendors leave out. Under the heading of its data boundaries, the documentation says Granola only has access to your meeting data or files that you upload or add to folders, that it doesn't know about your emails, offline tasks, or personal to-do lists, and that chat can therefore only answer questions based on what was discussed or documented in your meetings.

Read that against the questions you actually want answered. Where did we land on the renewal terms. Who owns the migration now. What did we promise them in writing. Each of those has an answer, and for most teams the answer is split across a call, a thread, and a document, with the deciding half living outside the meeting. A tool bounded by meeting audio can only ever return the part that was said out loud.

Fireflies has gone furthest in the other direction. Global AskFred is documented as searching past meetings, analysing information from connected apps including Gmail, Slack, HubSpot, Notion, Google Drive and Jira, running web search, and accepting custom MCP servers for tools not on that list. If breadth beyond meetings is your binding constraint, that is the honest answer on this shortlist, and it is worth knowing that connector access means the assistant can also act in those tools rather than only read them.

Otter has announced the same direction and not yet shipped it. AI Chat with Gmail, Google Drive, Notion and Jira appears on its pricing page under a Coming Soon heading, which is where it also sat when we last read that page a month ago. Fathom's integrations are built to push meeting content outward, into CRM records and into ChatGPT and Claude, rather than to pull your other work in.

This is the axis to decide first, because it is the one you cannot fix later with a plan upgrade. Every other limit here is a number on a pricing page. This one is what the product is.

• Write down your five real questions first, then check how many of them have answers that live outside a meeting.
• If most of them do, a meetings-only tool is the wrong category rather than the wrong tier.
• Treat a Coming Soon connector list as absent, and check the date on it.

Five questions to put to any vendor

These are answerable from documentation in about fifteen minutes per tool, and every one of them has a different answer across the four tools above. Send them to a sales contact and the speed of the reply tells you something too.

The reason to ask in writing is that each of these is a number or a boundary, not a judgement. You are not asking whether the assistant is good. You are asking what it reads, and no demo will surface that on its own, because a demo is always run inside a fresh account on the top plan with every meeting present and the transcripts intact.

• On the plan we would actually buy, how far back can a question reach, and is that measured in days, meetings, or stored minutes?
• If we shorten transcript retention for privacy reasons, which answers get worse, and does the interface tell us when we are asking about a meeting whose detail is gone?
• How many questions per month and per meeting does our plan include, and what happens when we run out?
• Does the scope of a question depend on where in the app we ask it, and is the scope shown in the answer?
• What can it read besides meetings today, with a link to the page that says so rather than a roadmap?

Where Driffle sits, limits included

Driffle was built on the far side of the last boundary. Meetings, documents, messages and the screen context from the apps you work in sit in one memory, so a question is not confined to what somebody said out loud on a call. That is the difference we would ask a buyer to test first, because it is the one that is structural rather than a line on a pricing page. Answers link back to where they came from, so you can open the moment behind a claim rather than taking the paragraph on trust.

On the depth boundary our position is simply stated: transcripts and notes are kept until you delete them, for as long as the account is active, so the detailed layer under an old note does not quietly thin out on a timer. Raw audio is the exception and is dropped within sixty seconds of transit, which is deliberate. You can edit or delete any segment, and a full export is available from settings.

The limits, since this article is an argument for reading them. Driffle is a Mac app for Apple Silicon, running macOS 14.4 or later, as a direct download. Intel Macs, Windows and Linux are a request form today, not a product. Transcription and the models that answer your questions run in the cloud, inside our own AWS account with no training rights granted to the model provider, so this is not an on-device or offline tool and we would rather say that plainly than let the word private imply it. And we are not exempt from the two boundaries that are numbers: our free tier limits how far back you can ask and how many questions you get, so put the first and third questions from the list above to us as well as to everybody else.

If the question you keep failing to answer is one whose evidence is scattered across a call, a document and a thread, that is the case Driffle was built for. You can download it at driffle.ai/download and ask it something from last month.