Where are the recordings stored?

Curious if they’re stored on device or in the cloud. If in the cloud, who has access?

3 Likes

@Atmosfan Recordings are stored locally. Pocket uses a “local-first” design with 128GB onboard storage. Everything is also end-to-end encrypted, and your data stays yours.

If and when it syncs to the cloud (for summaries, transcripts, etc.), it’s still encrypted. No one - not even us - can access your raw recordings.

9 Likes

I suppose the AI isn’t local though? Cloud based models running on the raw transcript?

@dps You’re right. Pocket doesn’t run AI locally. The raw audio is stored locally on-device first (privacy-first approach), but once uploaded, the heavy lifting, like summaries, action items, mind maps, runs in the cloud using top models (GPT-4, Claude, Gemini, etc.).

So:

  • Your data stays yours (encrypted, local-first by default).

  • The AI happens in the cloud, once you upload. That’s how it pulls off that level of insight.

4 Likes

@Naor Great questions. Here’s where things stand:

1. Can you bring your own model / change the transcription base URL?
Not at launch. The pipeline is tightly integrated with Pocket’s own backend to ensure quality, speed, and reliability - especially across languages, diarization, and action item extraction. That said, we’re watching what folks like you want. If enough people want to self-host or plug in custom endpoints, we might open it up.

2. Will Pocket be open sourced, especially the heavy lifting parts?
Parts, yes. But not the full stack. The core product relies on a lot of infrastructure: multi-model routing, hardware integration, local caching, secure syncing, and more. Some local-first elements may open up over time, but the full magic - things like AI diarization, summaries, maps - still rely on our stack.

That said, you keep your data, always. It’s encrypted and local-first by design.

3 Likes

With pocket using ChatGPT, Gemini, etc… for summaries, mind maps and other tasks will it be using a “closed community” meaning the work the AI is doing is not being utilized to train the larger model. When recording and using AI in some instances like if working in government or other industries if the recording that is then being transcribed, summarized, etc using AI if it then is being used to train the larger model it could be considered a public and therefore FOIA-able. If it is in a “subscription” based model wherein the information and data doesn’t trained the larger model then more than likely it is not.

4 Likes

Just a second for this request. I wouild love to be able to connect to my self-hosted models running in our private cloud using an API key. Let me know when you are ready to beta test that!

3 Likes

@Vickens8 everything you record with Pocket stays yours. Always.

Your data is not used to train the larger models. The AI models Pocket uses (whether GPT-4, Claude, or Gemini) are accessed through private, enterprise-grade APIs - the kind designed specifically for privacy, where user data isn’t retained or used for training.

9 Likes

Thank you for the super fast response!!

2 Likes

The data security layer will be opensource soon.

2 Likes

I would love to see some sort of open source solution for getting the audio recordings off the device, maybe with a basic whisper transcription feature.

3 Likes

We could do a MCP API if enough people upvote at feedback.heypocket.com

2 Likes

Hmm, that could be cool, but I want to see a way for us to get our recordings off the device without going through your cloud solution. I tried plugging into computer via USB-C and the computer didn’t see a volume at all, so not sure if that’s a viable way to access the files?

2 Likes

If enough people upvote and if something is possible, we will build anything.

The USB-C is only for charging

2 Likes

Thanks @Akshay and @dps ! I created feedback for this here: Transfer files to pc/mac via usb cable (cable is currently only for charging)

PLEASE UPVOTE THIS!

ps. I created a new Topic here for visibility: Transfer files to pc/mac via usb cable (cable is currently only for charging)

Is this a physical limitation or just all the FW allows right now?

2 Likes