Local-first as a hard constraint: who else is fighting the cloud-default expectation?
by•
Saw Hayley's stance on on-device transcription and patterns staying on the phone. We made the same call and it is harder than it looks once you say it out loud.
Minimi is an ambient memory layer for Claude. It captures what you see and what is said on calls, on your Mac, encrypted on-device. Keeping capture and storage local meant giving up some easy wins (cross-device sync, cheaper retrieval) in exchange for a promise we could actually stand behind.
Two things I am still wrestling with: retrieval quality without cloud-side processing, and how to explain local-first to people who assume cloud by default. How are you framing it?
We launched this week. If the local-first angle is your thing, I would genuinely like your read on it.
14 views
Replies
Shram
Quite a challenge.
Shram
Storage genuinely can be local. Creation is the catch. The second you need a model to build a memory out of raw capture, that step runs off-device today, because on-device models aren't good enough yet for the quality bar. So in practice local-first means local storage plus cloud processing with no retention.
Hayley: Your Thinking Companion
Really glad you asked this publicly because it forced me to get precise.
Here's exactly how Hayley works:
iOS: Audio never leaves your phone. We use Apple's SFSpeechRecognizer for fully on-device transcription. The text transcript is then sent to Claude API for reflection generation, that's the one cloud touch. We have no backend, no database, no logs of our own. All patterns and session history are stored in SwiftData on-device. Delete the app and everything is gone.
Android (launching this week in beta): Different story and worth being upfront about. Android uses Google's built-in Speech Recognition API, which means audio is processed by Google's servers for transcription. We don't store anything and have no backend, but it's a different privacy profile to iOS and I won't pretend otherwise.
The honest framing for both is exactly what Ojasvika said: local storage plus cloud processing with no retention. That's the defensible position we're comfortable standing behind publicly.
On your two open questions, retrieval quality is the harder long-term problem, agreed. On explaining local-first to cloud-default users, we've found "your voice never leaves your phone" lands better than any technical explanation for iOS. People understand audio privacy intuitively even if they don't understand inference.