Cairn
Leave a voice note. Come back to it later.
Create a place to speak, keep a moment, and revisit it without confusing a person’s original words with a generated reply.
- 48
- commits in three weeks
- 2
- devices: iPhone and Apple Watch
- On-device
- AI, speech and voice in the release build
- 17k
- lines of Swift
As of 2026-09-08, from the repository and releases


Captured from the development build
How it works
- 01
Leave a recording
User words and audio enter a session with explicit speaker attribution.
- 02
Keep the memory
The app stores the recorded moment and related conversation rows.
- 03
Return to it
Past-self prompts select and attribute context around the chosen moment.
- 04
Begin a dialogue
Generated words are kept distinct from the original recording and its author.
The hard parts, and what I did
Keep the moment. Leave room to return.
A voice note holds a person’s own words. A later reflection should help revisit it without replacing that original.
A journal that answers without the cloud
- The problem
- A voice journal holds the most private things a person says. Sending it to a server to be reflected on defeats the point.
- What I did
- In the release build, speech recognition, the reflection and the reply voice run on-device through Apple’s own frameworks. Remote models exist only as an opt-in tester setting.
- Result
- Native dependencies are fetched by pinned version and checksum-verified.
Keep the words, not just the summary
- The problem
- AI summaries flatten how someone actually said something, which is often the part worth coming back to.
- What I did
- Original recordings and transcripts are kept. A reflection is always tied to one past moment, so it can be checked against what was really said.
- Result
- The docs state plainly what has not been shown yet: no clinical or long-term claims.
Independent projectIndependent voice-memory project; generated reflections remain distinct from original recordings.
Stack
SwiftUI · ObservationSwiftDataAVFoundation · SpeechFoundationModelsOptional remote LLMTime-bounded recallKeychain configuration
Implementation details
Generated reflections are new responses, not recovered words or a recovered past voice.
- Your recording
- Keep the person’s original contribution distinct.
- Memory space
- Move between recorded moments and conversation.
- Conversation session
- Coordinate user and generated turns without mixing authorship.
- Past-self context
- Use a selected time perspective to frame a generated response.
- Local memory
- Store recording metadata and the associated audio files.
- Model route
- Use supported on-device generation or an explicitly configured remote route.