Voice Memos

Voice Memos

Memos recorded on iPhone or Apple Watch sync to the Mac through iCloud. scripts/memos.py reads them from the sync folder and transcribes them locally. Resolve it relative to this SKILL.md, following the skill folder symlink to its source, and call it by absolute path:

python3 "<resolved-skill-directory>/scripts/memos.py"

• Without options it takes the memos added since the last run. The first run takes only the newest memo.
• --last 3 takes the newest three, handled before or not.
• --since 2026-09-01 takes memos from that day on.
• --dry-run lists what it would take, without transcribing.
• --language ja is only needed for languages outside Parakeet's 25 European ones. It switches to Whisper.

Expect about a minute per hour of audio. The first transcription also downloads the model (2.4 GB), so give the command a long timeout.

Requirements

A Mac with Apple silicon, brew install ffmpeg uv, and Voice Memos turned on under iCloud on the Mac and the phone.

macOS guards the sync folder. When the script says it is blocked, tell the user to add the terminal app that runs the agent (Ghostty, Terminal, iTerm) under System Settings > Privacy & Security > Full Disk Access, restart it, and run the skill again. Do not look for a way around the block.

After transcribing

1. Read each whole transcript file.
2. For each memo, give the recording time, then the text. Summarize a long memo when the user did not ask for the full text.
3. Then do what the user asked. When they want a message out of it, hand the text to the matching skill, for example polish-text.

Traps

• Transcripts garble names. Keep a name as transcribed and flag it unless the user confirms the spelling. Do not pick between similar names from notes or memory.
• Memos are private. Do not paste them into other tools or files unless the user asks.