Building a personal knowledge base from meeting notes

Niklas Luhmann's Zettelkasten is the origin story: roughly 90,000 index cards, from which he produced more than 70 books. The system's power was never storage. It was two rules — one idea per card, and every new card linked to the cards it related to — which meant that looking up one thing surfaced its neighbours, including the connections he had forgotten making. He described the box as a communication partner rather than an archive.

Three ideas from that tradition do the work here:

• Atomicity (Zettelkasten, and Andy Matuschak's evergreen notes) — one note holds one idea, which is what makes it linkable from more than one place. A note containing six things can only be filed under one of them.
• Organize by actionability (Tiago Forte's PARA: Projects, Areas, Resources, Archives) — sorting by what something is for rather than what it is about, which is what stops filing turning into taxonomy design.
• Maps of Content (Nick Milo's LYT) — a note that is a curated list of links, made when you have too many notes on a theme to hold in mind. Structure emerges from the material instead of being imposed before it exists.

All three were developed for material you read. Meetings are a different input — conversational, redundant, and generated at a rate no one can process by hand — and that difference is the whole of what follows.

The most common failure is treating the transcript as the knowledge base. It is not, for the same reason a stack of photocopied articles is not a Zettelkasten: it is unprocessed. A transcript is verbatim, full of false starts, and organized by time — the one axis along which you will never search.

Karpathy's LLM Wiki pattern names the layers cleanly: raw/ holds the immutable sources, and a compiled layer holds what was extracted from them. For meetings the mapping is exact — transcripts are raw/, permanent and never edited, and the knowledge base is what you distil from them. The transcript's job is to be the source you can always go back to when the note turns out to be wrong or incomplete, which is a job it does better than any note-taker.

What crosses from one layer to the other is a small fraction. A one-hour meeting typically yields between one and three permanent notes, and most meetings yield none — which is fine and worth saying out loud, because the instinct to extract something from every meeting is what fills a knowledge base with noise.

The atomicity rule translates well: one decision, one note. Not one meeting, one note. A meeting that decided three things produces three notes, each linkable from a different project, and that is precisely the property a dated meeting summary lacks.

Four kinds of note come out of meetings, and they behave differently:

Write each in your own words, with a title that states the claim rather than the topic: "Enterprise customers block on SSO before evaluating features" rather than "SSO." Matuschak's argument for this is that a title stating a claim forces you to have understood it, and it makes the note findable by what it says rather than by what it is filed under. Link back to the transcript, which is what lets a compressed note stay trustworthy.

PARA's answer — organize by actionability — handles most meeting output without argument. A decision about the Q4 launch goes in that project; a constraint about how procurement works belongs to an area, since it outlives every project it touches; a concept goes to resources; and when the project finishes, the whole folder moves to archives, which is the step that keeps the active space small.

The mistake worth avoiding is building the hierarchy first. Forte's own advice is that structure should follow use, and Milo's Mental Squeeze Point makes the same point from the other side: create a Map of Content when you have accumulated enough notes on a theme that you can no longer hold them in mind, not before. A folder tree designed in advance is a prediction about material you have not gathered yet, and it will be wrong.

Chronological notes answer "what happened on the 14th." A linked knowledge base answers the questions that matter: why does this constraint exist, when did our position on this change, what did the customer say the first three times this came up. Those questions span months and are unanswerable from dated files, because you would have to know which date to look at.

Three linking habits produce most of the value:

• Link the supersession. When a decision replaces an earlier one, link them and say what changed. The pair is more informative than either note alone, and it is how a knowledge base records that it learned something.
• Link the recurrence. When the same concern appears in a third meeting, link the three. Three unconnected mentions is noise; three linked mentions is a pattern, and the linking is what converts one into the other.
• Make a Map of Content per project retrospective. At the end of a project, one note listing the decisions in order with a sentence on how each turned out. This is the single highest-value artifact in the whole system, and it takes an hour if the notes exist.

Every method above assumes reliable raw material. This is where meeting-based PKM historically fails: the extraction has to happen from memory, days later, because nothing complete was captured at the time. You end up distilling your own summary of a summary, and the constraint someone mentioned at minute 40 — exactly the thing that would have been valuable in March — was never written down at all.

Earkeep supplies the raw/ layer. It records mic and system audio continuously while it runs and transcribes locally on your device, so every meeting leaves a complete record without a note-taker — no bot to invite, no start/stop to remember. Meetings are defined afterwards by selecting a span on a timeline, or automatically from a connected calendar. Audio is processed only in memory and never written to disk, and each day is a plain .jsonl file in a folder you choose, so an Obsidian vault, a git repo, or a shell script can all read it. That last property is the one that matters over a decade — a knowledge base whose value compounds cannot live in a format that expires with a subscription. See plain files beat platform lock-in.

An agent can do the first pass of distillation: read the day's transcripts, propose the decisions and constraints, suggest which existing notes they connect to. What it should not do is write the note for you unsupervised. The PKM tradition is unanimous that the value is in the thinking — Matuschak's "better thinking, not better note-taking" — and a knowledge base of unreviewed AI summaries is a search index with extra steps. Use it to propose; keep the judgment. The mechanics are in feeding meeting context to AI agents and local-first AI at work.

Download Earkeep free for macOS and stop distilling meetings from memory.