Where AI Notetakers Store Your Meetings (2026)
On this page
Short answer - where does each one keep your meetings?
- EU by default: tl;dv (Germany and Finland).
- US by default, EU on a sales call: Fireflies (Enterprise, and even then it's stored in the EU but processed in the US) and Read.ai (Enterprise+).
- US, no EU option on their pages: Otter, Granola, Fathom.
- Your own Mac: Humla, which is the one I make. More on that, and on why I'm not a neutral party, below.
Every row here comes from the vendor's own privacy policy, security page, help guide or pricing page, read on 17 November 2026. Where a page didn't say something, the table says "not stated" rather than me filling the gap from a third-party blog or my memory.
Why I went looking
Germany is the second-biggest country in this site's analytics, right behind the US, and the UK is third. So a lot of you reading this are in Europe, and "where is my client call actually sitting right now" is not an abstract question for you. It's a GDPR question, and sometimes a question your client's procurement team asks before they let you record anything.
I assumed the answer would be boring: "everyone's on AWS in Virginia, done." It mostly is. But the interesting stuff lives in the footnotes (it always does), and the footnotes turned out to disagree with each other in ways the marketing pages don't.
Bias, up front, because it's going to be sitting next to me the whole post: I make Humla, a free and open-source Mac app that keeps meetings on your own disk. I have an obvious interest in you finding the cloud answers uncomfortable. So I've tried to give credit where it's due, and two of these vendors have earned some.
The table
| Default storage | EU option | Default retention | Trains on your meetings? | |
|---|---|---|---|---|
| Otter | US (AWS) | Not stated | No fixed period; deleted conversations purged from Trash after 30 days | Yes, de-identified; no opt-out described |
| Fireflies | US (AWS and GCP) | Enterprise "Private Storage": stored in the EU, processed in the US | Not stated; account data deleted within 30 days of closing | No, per its privacy policy |
| Granola | US (AWS) | Not stated | Notes and transcripts kept indefinitely; audio not kept after transcription | Yes, anonymised; opt out in Settings |
| Fathom | US | Not stated | No fixed period; custom retention on Enterprise, by email | Yes, de-identified; opt out in settings |
| tl;dv | EEA (Germany, Finland) | It's the default | Free: 3 months; paid: until you delete the account | No, per its privacy policy |
| Read.ai | US and other countries | Enterprise+ ("EU & CA hosting") | Audio and video at most 2 years; paid: until you stop paying or delete | Opt-in programme for meeting data, plus model improvement "consistent with your account settings" |
| Humla | Your Mac | Optional Cloud sync is hosted in the EU | Until you delete it; audio not kept by default | Nothing to train: Humla has no model of its own |
Sources, all checked 17 November 2026: Otter's privacy policy and privacy & security FAQ; Fireflies' privacy policy and data storage guide; Granola's security page and privacy policy; Fathom's privacy policy; tl;dv's privacy policy; Read.ai's privacy policy and pricing. Policies change. Check before you rely on a row.
Storage region: five Americans and a German
This is the column most people came for, and it's the least surprising. Otter, Fireflies, Granola, Fathom and Read.ai all keep your meetings in the US by default. Fathom's privacy policy is the bluntest about it: the service is hosted in the United States and intended for visitors there, and European users are, in effect, told they're choosing to send their data across.
The EU options are where it gets fiddly:
- Fireflies sells Private Storage on its Enterprise plan, and its own guide is refreshingly honest about the catch: your data is stored in the EU, but processed in the US. So the recording rests in Europe and takes a round trip to America every time something happens to it. Whether that satisfies your client's DPA is a question for your client's lawyer, not for me.
- Read.ai lists "EU & CA hosting" as an Enterprise+ feature, which needs five or more licences and a sales conversation.
- Otter, Granola and Fathom don't mention an EU option on any page I read. That's not proof one doesn't exist behind a sales call. It does mean you can't buy it with a credit card.
And then tl;dv, which just... does it. Its privacy policy says everything is hosted in the European Economic Area (on Google Cloud, Hetzner and Wasabi, with data centres in Germany and Finland named), with one carve-out: the AI part runs in the EU or the US depending on the hosting location you pick in your account. (The company behind it is tldx Solutions GmbH, and the EU-first setup shows.)
Retention: "as long as necessary" is not a number
Here's a phrase you'll see a lot: "as long as necessary to carry out the purposes for which we collected it." That's Otter, Fathom, Fireflies and Read.ai, more or less word for word. It's a legal phrase, and what it means in practice is "until you delete it, and we'll decide what necessary means."
The vendors that actually give you a number:
- Granola keeps notes and transcripts indefinitely unless you delete them, offers custom retention on Enterprise, and removes everything within 30 days of account deletion. Credit where it's due: its privacy policy says it doesn't keep the audio once the transcript exists. That's a real privacy choice and most of the others don't make it.
- tl;dv keeps free-plan recordings and transcripts for 3 months, and paid ones until you delete the account.
- Read.ai caps audio and video at 2 years, and keeps paid accounts' data until you stop paying or ask for deletion. Custom retention is Enterprise+ only.
- Otter empties its Trash after 30 days. Delete a conversation and it's really gone a month later (or immediately, if you clear the Trash yourself).
- Fathom promises "commercially reasonable efforts" to delete within 30 days of an account deletion request, and custom retention is Enterprise only, by emailing support.
Notice the pattern? On almost every one of these, automatic deletion is an enterprise feature. If you're a freelancer or a five-person agency on a paid plan, the default retention policy is: forever, or until you remember.
Training: where the footnotes disagree
This is the column where the fine print matters most.
Fireflies and tl;dv say no. Fireflies' privacy policy says it doesn't use personal information for AI training, and that meeting content isn't used to train internal or external models. tl;dv's says it doesn't use customer content (recordings, transcripts, notes) to train or improve models, for itself or anyone else. Two clean answers. I'll happily admit I didn't expect Fireflies to be one of them.
Granola and Fathom say yes, with an off switch. Both train in-house models on de-identified data, and both let you opt out in your account settings. Granola's Enterprise workspaces are opted out by default. Neither page says the opt-out is limited to a plan.
Otter says yes, and doesn't mention an off switch. Its privacy FAQ says it de-identifies user data before training its models, and that the process is automatic. I couldn't find an opt-out for individual users on its privacy policy or its FAQ. (Its help centre blocked my requests, so if one exists there, I haven't seen it.)
Read.ai is the hardest to summarise. The policy describes a Customer Experience Program you opt in to, which lets meeting information and connected email and messages be used to improve features. Separately, it lists improving and training "our models within our Services" among the things it does, "consistent with your account settings." Read the Read.ai policy yourself if this one matters to you. I'm not going to paraphrase a double negative into a promise.
One thing all of them agree on: the big AI providers underneath aren't allowed to train on your data. Granola names OpenAI and Anthropic, Fathom names OpenAI, Anthropic and Google, Otter says its AI providers don't train on customer data, Fireflies says its vendors are contractually barred, and tl;dv says the same of its providers. That's the API-terms default these days (OpenAI's and Anthropic's both say so), so it's good hygiene rather than a selling point.
And Humla, since Bias is still sitting here
Fine. Here's my row, checked against Humla's own source on the same day.
By default, Humla doesn't put your meetings anywhere except your Mac: a SQLite database in your user Library. The audio is thrown away after transcription unless you turn on Keep audio in settings, and even then it stays on your disk. There's no retention policy because there's no server to retain anything on. You delete a note, it goes to the Trash, you empty the Trash, it's gone.
Two honest asterisks:
- If you pick a cloud transcription or AI provider, that provider's terms apply to whatever you send it. I've written up exactly what OpenAI, Deepgram, Groq and Anthropic keep. Or skip them: transcription and summaries can both run on-device.
- If you turn on Humla Cloud to share notes with a team, those notes sync through our servers, which are hosted in the EU. That's opt-in, it's $5 per user per month, and it's explained on the Teams page.
And training: Humla has no model of its own, so there's nothing to train. That's not a virtue, it's just architecture. But it is the kind of architecture that makes this whole post unnecessary.
So which one?
If you need a cloud notetaker and you're in Europe, tl;dv is the only one here that's EU-hosted without a sales call, and it says it doesn't train on your meetings. That's a strong combination and I'd be lying if I pretended otherwise. (Here's how it compares to Humla, for the full picture.)
If training is your main worry and residency isn't, Fireflies' policy is as clean as tl;dv's. Granola and Fathom are fine once you've found the toggle, so go find it today. Otter is the one I'd ask questions of before recording anything sensitive.
If the real answer is "I'd rather nobody stored it at all," that's the local-first option, and it's what I built Humla for. It's free and open source, it never joins the call as a bot, and the region your meetings are stored in is "the laptop in front of you." Bias is nodding. For once, it's allowed to.
Michael Wilhelmsen builds Humla, the open source meeting notes app for Mac, from Norway.