249 Voice Memos: What Was Inside

I had 249 voice memos on my phone.

Most of them were untitled. Some went back to 2017. Every “I should write that down” moment I’d ever had, recorded and then immediately lost.

Last month I decided to actually do something with them.

I built a skill file that transcribes all of them locally on my mac using whisper.cpp, no API keys, nothing sent anywhere, then runs each transcript through a local model to generate a title, a summary, and a list of themes. The output is a searchable index: one entry per memo, organized by theme, with anything that looks like a story or a distinct insight flagged for further use.

The whole run went overnight. When I woke up I had a searchable index of nine years of my own thinking.

58.8 hours of audio. 68 of those memos became distinct atoms in my content library.

Including the actual origin story of Helm. I remembered recording it but had no idea where it was. It’s a 1-minute-40-second voice memo from July 2025 where I’m describing the moment I realized no one had built the right focus tool for ADHD brains. Sitting on my phone the whole time.

There was also an entire brand narrative I recorded in early 2025 and completely forgot. It’s basically the thesis for what Little Might is. I recorded it in a parking lot apparently.

---

Here’s the thing I had wrong for nine years: I thought I had a capture problem.

I blamed myself for not being more disciplined about journaling, more consistent about taking notes. But I was capturing things. Constantly. The problem was retrieval. My phone captured. My brain captured. My notes app captured. None of them surfaced anything when I actually needed it.

What AI changed isn’t the ability to store things. It’s the ability to summarize and index things well enough that you can finally use what you already have.

I didn’t need a new second brain. I needed to process the one I’d already been building for nine years without knowing it.

---

If you have an ADHD brain and a backlog of untitled voice memos, this is worth a weekend.

The short version of the setup:

• install whisper.cpp locally (brew install whisper-cpp) • find your voice memos in Finder (they sync from iCloud to ~/Library/Application Support/com.apple.voicememos/) • run whisper on each file to generate transcripts • pipe each transcript to Claude with a prompt: “give me a title, a 2-sentence summary, and 3-5 themes. flag if this contains a distinct story or insight worth developing.” • build an index from the outputs

I’ll share the full skill file in a future issue once it’s cleaned up enough for other people to use without reading my internal agent docs.

---

The question worth asking isn’t “what should I be capturing?”

It’s “what have I already captured that I can’t find?”

---