From meetings I keep only what still holds in six months

I record almost every meeting I can, and I already have dozens of transcripts. A summary of what was discussed and who said what isn’t that important in my view, because a week later nobody reads it anyway. Tasks fall out of the transcript either way, so the agent just proposes them and I confirm.

What interests me more is what from the meeting will still hold in six months. For example a decision together with its reason, an agreed way of doing something, or a pattern that keeps coming back across meetings.

Most meetings leave me with just a few tasks, because it’s regular operations, and some leave nothing. When there’s nothing more in the transcript, the agent shouldn’t force anything out of it, so a regular status gets one line or gets skipped.

The things that hold longer go into a memory layer that both I and my agent work with. I’ll write about that layer separately, because it’s a longer story, and here I’ll show how things get from a meeting into memory.

How I have it set up

I record in MacWhisper, because of everything I tried, it’s the one I trust most. I don’t want anything else from it, just the recording. The text is then written by Soniox (stt-async-v5), which I have plugged into MacWhisper in place of Deepgram. Who said what is decided by NVIDIA Nemotron 3 Diarization. It runs locally on my Mac through NeMo-Speech.cpp, about 3.7 seconds per audio minute on an M2 Pro, and for free. If it fails, pyannoteAI in the cloud takes over.

With Nemotron I got burned by the default setting. The offline preset merged two of four people into one in a 108-minute meeting, while the streaming preset matched pyannote on about 98% of the recording. So test the default settings on your own recordings first.

Names turned out to matter more than the models themselves, because a transcript where everyone is “Speaker 2” is useless three months later. I used to fix them directly in MacWhisper. Now every transcript lands in Command OS, the app I built to run the company, and I fix names there. I name a speaker once and the name sticks for every later meeting, so I can then ask, for example, for every meeting where two specific people talked.

The processing runs in pi, an agent harness. The agent reads the transcript in chunks, checks names against my list of people, and writes the result back to Command OS. It only proposes tasks and things for memory, so nothing gets in until I approve it. For the weeks when I don’t get to it, I have a background job. After 72 hours it writes a short TL;DR and links the people for every meeting that stayed unprocessed.

The prompt

The core of the prompt is below, and I trimmed it so it works without my setup. The most important part is the question “what from this will be useful in 3 to 6 months?”, because the agent uses it to decide what to keep and what to throw away.

You process a meeting transcript. Don't optimize for a faithful summary.
Optimize for future usefulness. For every part, ask:
"What from this will be useful in 3-6 months?"

1. Verify names. Speech recognition breaks names. Check every person and
   company against the list I give you. If unsure, ask me.
2. Split the content into three layers:
   - Evergreen: decisions, recurring patterns, role clarity, lessons.
   - Operational: follow-ups, blockers, prep for the next meeting.
   - Noise: small talk, debate details, who-said-what chronology. Drop it.
3. Pick the smallest output that fits:
   full note / short note / one-line stub / nothing.
   Don't invent depth for an operational meeting.
4. Note format: TL;DR (max 6 bullets, written to me), Key insights,
   Decisions, Open questions, Action items.
5. Propose tasks and long-term memories. Never create them. Wait for my "yes".
6. Put each insight in one place only. No copies across note, tasks and memory.