How it works
How it stays yours.
Stonekept lives across your iPhone, your Mac, and if you want it, Claude Desktop. Here is what actually happens to a recording at every step. In plain terms. No marketing fog. No claims you can't verify.
It starts with one tap.
Tap your Apple Watch, click the menu bar on your Mac, or open the iPhone app. A Live Activity on the iPhone lock screen shows the mic is on, the entire time it's on. No wake word, no background listening. The mic stays off until you tap record.
Sessions can be a thought, a conversation, or a whole afternoon. End the session yourself, or let it stop on its own. Either way, the audio sits on the device, encrypted at rest, until the next step finishes.
Capture is the only step that ever touches a microphone.
Or upload audio from anywhere else.
You don't have to record through Stonekept. Drop an audio file in from your iPhone or Mac and it picks up at the next step. A Voice Memo, audio from a handheld recorder, an audio-only meeting export, an old interview on your hard drive. We accept m4a, mp3, wav, aiff, flac, and opus.
Same on-device path as a live session. Transcription runs on your device. Extraction runs on your device. The original file sits in your library next to the rest of your memories.
Your recorder, your file, our pipeline. Still on your device.
Your phone reads it.
The moment a session ends, your iPhone transcribes the audio on the iPhone itself. Apple's built-in speech model runs first. If it stumbles in a noisy room or on accented speech, a backup model runs as well, also on the device.
There is no upload step here, even for Pro users. The transcript is what feeds the next step. The audio stays put until extraction is done with it.
Your voice never reaches our servers.
The parts worth remembering, lifted out.
A recording is just a recording. Nobody listens back to a 47-minute file. Stonekept's AI reads the transcript and pulls out the things your brain would have wanted to keep. Decisions. Tasks. People. Patterns. Ideas. Each one becomes a structured memory you can search, link, and come back to.
This AI lives on your device. It downloads once, runs on the Neural Engine, and never calls home. Extraction is the heaviest part of the process. Long sessions are best processed when the phone is locked and charging, which usually means overnight.
Once extraction finishes, the audio is discarded by default. What remains is the structured memory, in your library, waiting for review.
On-device AI. No cloud. No model API. The phone does the work.
Nothing lands in your library without your yes.
Every extracted memory shows up in a queue. Keep what matters. Edit a fact. Toss the rest. You see the source quote, how confident the model was, and the conversation it came from.
Review on your iPhone over coffee, or open the Mac app for the same queue with more breathing room. Either way, your library only contains what you've signed off on.
The AI drafts. You decide.
Sync without seeing inside.
Add a second device, usually a Mac, and your memories follow you. Each memory is locked on one device with a key that only your devices have. The locked package travels through our servers, but our servers cannot open it. We carry envelopes. We never read the letters.
Lose your recovery phrase and we cannot restore your library for you. That is the cost of actual privacy. The recovery phrase is set up when you turn on sync, and you only need it if every device is gone.
Encrypted in transit and at rest. End to end. Real.
Ask your memories anything, on your device.
Open Chat in Stonekept and talk to your own life. Ask what you decided about the v2 launch. Ask what someone said about Tuesday. The model that answers runs on your iPhone or Mac. Nothing leaves the device. No sign-in, no cloud, no catch.
Same on-device AI that does extraction. Different prompt.
Bring your memory to Claude.
On Pro, you pick which memories Claude can read. Stonekept hands them over only when you ask, only the ones you chose, and only for that conversation. Anything else stays in your library. There are two ways the connection runs, and you choose which one is on.
Local-to-local on Mac. Claude Desktop on your machine talks to Stonekept on your machine through a local channel. Nothing leaves the computer. Anthropic sees what you ask Claude (it is their app), but we never see the question or the answer.
Cloud relay on phone or claude.ai. When you are not at your Mac, Stonekept's broker carries the chosen memories between Claude and your library. The broker is a relay, not a database. Nothing is stored. Nothing is logged. Nothing is analyzed. Traffic is encrypted on the wire and the broker forgets every request the moment it is done.
Stonekept never stores or analyzes your memories. Claude only sees what you choose. Once a memory is in a Claude conversation, Anthropic's normal terms apply to that conversation. You can turn the connection off any time.
A note on model sizes
Different devices run different sized models.
Stonekept Extract comes in two sizes: a 4 billion parameter version and a smaller 2 billion parameter version. The one your device runs depends on how much memory it has. Same things come out of every tier (decisions, tasks, people, ideas), but the bigger the model, the longer the verbatim quotes and the richer the detail.
Below 6 GB of RAM, extraction runs on a paired Mac instead of the phone. Capture itself works on any supported device.