Turn meeting transcripts into an AI-ready knowledge base

A meeting transcript is one of the richest inputs you own and one of the least useful. Everything that mattered is in there: the decision, who owns the follow-up, the number someone quoted. It is just buried under two thousand words of "yeah, totally" and "can you hear me now."

Recording and transcription got solved. Every call tool ships a transcript now. The unsolved part is what happens next. The transcript lands in a folder, nobody reads it again, and a week later someone asks "wait, what did we decide about pricing?" The knowledge was captured. It just was not made usable.

This post is about closing that gap: taking a raw transcript and turning it into structured, AI-ready context that you and your AI tools can actually retrieve. Not a summary you skim once, but durable context that compounds.

What "AI-ready" actually means

An AI-ready note is not a prettier transcript. It is a note whose important facts have been lifted out of the prose and into a structure a machine can read at a glance. In practice that means a small block of metadata at the top of the file:

---
type: transcript
participants: [Maya Chen, Devin Okafor]
projects: [Voxbit]
decisions:
  - Ship the v1 beta on Friday
  - Cut team workspaces from launch scope
action_items:
  - Maya to finalize onboarding copy
  - Devin to migrate the search index
summary: Locked the beta date and trimmed scope to hit it.
---

[00:01] Maya: okay so the real question is the date...
(the full transcript stays here, untouched)

The body of the transcript does not change. You keep the full record. What you add is a layer on top: the same information, structured. Now a question like "what has Devin committed to across every call this quarter?" is answerable, because action_items is a field, not a sentence hiding on line 340.

Why frontmatter, and why markdown

You could put this structure anywhere: a database, a Notion property, a spreadsheet. The reason to put it in YAML frontmatter inside a plain markdown file is that it survives. It is readable with the AI turned off. It is portable to any editor. It is not trapped inside one vendor's format, which matters a great deal the moment you want a different AI model, or a different tool, to read it.

This is the difference between a summary and a context layer. A summary is a one-time artifact. A context layer is a structured, living representation of what you know, kept in files you own, that any tool can query. Transcripts are the single highest-value input you can feed into one, because a meeting is dense with exactly the facts that structure loves: people, decisions, commitments, dates.

The extraction, step by step

Whether you do this by hand, with a script, or with a tool that does it for you, the shape is the same:

1. Detect that it is a transcript. Filenames help (call, meeting, transcript), and so does the shape of the text: repeated speaker labels, timestamps.
2. Pull the participants. Names attached to speaker turns. This is what lets you later ask "every call with the design team."
3. Pull the decisions. The statements that closed a question. These are the highest-value lines in any meeting and the easiest to lose.
4. Pull the action items, with owners. "Devin to migrate the search index" beats "someone should look at search." An owner turns a note into a system that can chase itself.
5. Write it to frontmatter. Leave the body alone.

The one instruction that matters most when you use an AI model for this: tell it not to invent. A model asked to "find the decisions" will happily manufacture one if the meeting was inconclusive. The extraction should pull what is there and leave the rest empty. An honest empty decisions field is worth more than a confident wrong one.

Where it compounds

One structured transcript is useful. The payoff is what happens across many. Once decisions and action items live in fields, you can roll them up. Every transcript that mentions Voxbit can feed a single project page: recent decisions, open loops, who is involved, all assembled from the meetings themselves. That page is not something you maintain. It is compiled from the structure you already extracted.

This is the quiet reason to bother. The transcript stops being a dead archive and becomes an input to a system that gets sharper the more you feed it. Ask your AI "what is the status of Voxbit?" and the answer is grounded in the actual decisions from the actual calls, not a vibe.

How LocalBrain does it

This is exactly the loop LocalBrain runs on your own files. Drop a transcript into your vault and its Autopilot detects it, extracts participants, decisions, and action items into the note's frontmatter, and leaves the body verbatim. Those fields then feed compiled context pages, one per project and person, that stay current on their own.

Because everything is plain markdown in a folder on your machine, none of it locks you in, and because every automated change is journaled and reversible, you can let it run without worrying it will quietly mangle a note. If an extraction is off, you revert it in one click. That combination, structure plus trust, is what makes it something you can actually leave on.

The transcript was never the problem. The gap between the transcript and the decision was. Close that, in files you own, and the meeting you barely remember becomes context you can query for years.