Apple Voice Memos

Apple Voice Memos

Extract and process transcripts from Apple Voice Memos synced via iCloud.

Prerequisites

Voice Memos must be synced with iCloud on macOS.

Tools

This skill includes two scripts in its scripts/ directory. Script paths in this document are relative to this skill's directory — your working directory is likely elsewhere, so invoke them with absolute paths (e.g. python3 <skill-dir>/scripts/extract-apple-voice-memos-metadata).

• extract-apple-voice-memos-metadata — Queries the CloudRecordings.db SQLite database (read-only) and outputs CSV with columns: title, date, duration, path, has_transcript. Supports optional flags: --limit N (default 10), --offset N, --search TERM, --after YYYY-MM-DD, --before YYYY-MM-DD.
• extract-apple-voice-memos-transcript — Extracts the embedded transcript from a .m4a file's tsrp atom. Outputs timestamped text with filler words removed, intelligent line breaks, and paragraph breaks at natural pauses.

Step 1: Select a voice memo

Run the metadata script to find the right recording. Choose flags based on what the user asked for:

• No specific request → run with no flags (returns 10 most recent)
• User mentions a topic or keyword → use --search TERM
• User mentions a time period → use --after YYYY-MM-DD and/or --before YYYY-MM-DD
• User wants to see more results → use --offset N to paginate, or --limit N to increase the batch size

Flags can be combined, e.g. --search work --after 2026-01-01 --limit 5.

python3 scripts/extract-apple-voice-memos-metadata [flags]

If the user's request unambiguously identifies a single memo (e.g., "my latest voice memo", or a search that returns exactly one match), proceed directly with that memo without asking. Otherwise, present the results as a numbered list showing title, date, and duration — mark any memo with has_transcript = no as "(no transcript)" — and ask the user which memo they'd like to work with.

Error handling:

• "Database not found" → Voice Memos iCloud sync is not enabled on this Mac.

Step 2: Extract and process the transcript

Spawn a subagent with fresh context to extract and process the transcript. Do not run the transcript script yourself first — letting the subagent extract it keeps the full transcript out of the main conversation context. The subagent prompt must contain:

1. The title and path values of the selected recording (from the metadata output)
2. The absolute paths to this skill's scripts/extract-apple-voice-memos-transcript script and PROMPT.md
3. These instructions: run the transcript script with the recording path, read PROMPT.md, append the memo title after the ## Memo Title heading and the transcript after the ## Transcript heading, then follow the complete prompt and return the resulting markdown document.

If your environment has no subagent or task tool, do the same work yourself: run the transcript script, read PROMPT.md, and follow its instructions directly with the extracted transcript.

Only run the transcript script directly if the user explicitly asks to see the raw timestamped transcript:

python3 scripts/extract-apple-voice-memos-transcript "<FILENAME>.m4a"

Error handling (reported by the subagent):

• "tsrp atom not found" → This recording does not have an embedded transcript. Apple generates transcripts on-device and not all recordings will have one.
• File not found → The recording file may not have synced to this Mac yet.

Step 3: Present the output

The subagent will produce a markdown document starting with a # <Memo Title> heading, followed by narrative summary, detailed notes, asides, and action items. Insert a metadata block directly beneath the title heading, then present the output to the user:

# <Memo Title>

- **Date:** <date from the metadata output>
- **Duration:** <duration from the metadata output>
- **Source:** <FILENAME>.m4a

After presenting the output, ask the user if they'd like to save it as a markdown file. Suggest a filename in the format YYYY-MM-DD-slugified-title.md derived from the memo's title and date (e.g., 2026-02-04-the-soul-of-a-new-machine.md). Save to the current working directory by default. The user may accept, provide a different name or path, request adjustments to the content first, or skip saving.