The half-second before you write it down
Your knowledge base is not a record of what happened. It's a record of what cleared your threshold for significance — and the threshold is invisible to every tool ever built.
There's a moment in every good conversation where something lands and you think: I should probably write that down. Then you don't. The meeting moves on, the thought dissolves, and your notes from that day reflect everything formal enough to commit to — not the thing that actually changed how you think.
Every knowledge tool ever built is designed for after that moment. None of them has any purchase on it.
The filter nobody sees
Your knowledge base is not a record of what you experienced. It's a record of what cleared your personal threshold for significance. That's not a small distinction. The threshold decides what makes it in — and everything under it disappears without trace.
What determines the threshold? The cost of capture, mostly. When writing something down takes ten seconds, you'll do it for things that feel worth ten seconds. When it takes thirty, you'll do it less. Note-taking researchers have found that people make rapid implicit judgments about whether each piece of information warrants the cognitive cost of transcription — judgments that happen below the level of deliberate thought. The threshold isn't a policy you set. It's a reflex conditioned by the friction of the tool.
Paper notebooks and digital apps produce different thresholds in practice, not because paper is slower (though it is) but because the friction of handwriting makes you decide whether to write before you start writing, not halfway through. On a laptop, you can start typing and see whether it was worth it after the fact. That's a different cognitive act, and it produces different notes.
Craik and Lockhart, working on memory encoding in 1972, found something relevant here: the depth at which you process information at the moment of encoding determines how well you remember it later. Making a semantic judgment — is this meaningful? does this connect to what I know? — does more for retention than simply transcribing. The act of deciding something is worth writing down is itself a form of processing. Skip the judgment, and you skip some of the encoding too.
What low friction buys you
The ambient recording tools — Otter, Zoom's AI notes, the wearable devices that transcribe everything you say — sidestep the threshold decision entirely. No judgment required. Every word gets captured.
The appeal is obvious. The thing you decided wasn't worth writing down at 2 PM might turn out to matter at 9 PM. If everything was captured, you'd have it.
But the threshold problem doesn't disappear. It moves. Instead of filtering at capture, you filter at retrieval. Which is worse, because now you're sifting a transcript of everything, with less context about what mattered, at the moment of need rather than the moment when the context was still alive. The ambient recording gives you a river instead of a bucket. The river contains more, but it's harder to drink from.
What gets lost in either approach is the same thing: the reasoning behind the filter. When you decide something is worth writing down, you know why. When an AI transcribes everything, no one does.
The question typed entities actually ask
When you write a note in a blank document, the question is implicit and unanswerable: Is this worth capturing? It's a probabilistic judgment about future need, made with incomplete information, under time pressure. No wonder people get it wrong in both directions — too much and too little.
Typed entities — person records, decision records, preference fields — change the question. They ask: What kind of thing is this? That's a different question, and a much easier one. Not "should I write this down?" but "where does this go?" The category does the filtering. If something fits into a person record, it goes in. If it doesn't fit any category, maybe it really wasn't worth capturing.
This sounds like a small shift. It isn't. The taxonomic question forces a moment of processing that the blank note never demands. You're not just transcribing — you're classifying. And classifying requires understanding what something is.
Luhmann noticed this about his Zettelkasten, which he developed over nearly forty years from the early 1950s until his death in 1998. The system forced him to ask, for every note: where does this connect? That question made every note a relationship, not just a record. The act of placing something changed how he understood it. He wrote 70 books. Whether he was unusually productive or unusually systematic is hard to separate from the methodology.
What your AI inherits
When you connect an AI to your knowledge base, it inherits everything that cleared your threshold. It also inherits — invisibly — everything the threshold excluded.
An AI reading your person records doesn't know what you noticed about someone and decided wasn't significant enough to write down. It doesn't know about the conversation where your opinion shifted, if you didn't capture the shift. It treats the written record as complete. The threshold is invisible to it, so the gaps are invisible too.
This isn't a flaw of AI tools specifically. It's a flaw of any system that treats captured knowledge as a faithful picture of what you know. The knowledge base is a projection of what happened through the lens of what you decided to record. Everything outside the lens is gone, and there's no metadata indicating anything was ever there.
The honest implication: the most trustworthy knowledge base isn't the one that captures the most. It's the one where the threshold was set thoughtfully — where what cleared it did so for reasons a future reader could reconstruct. Typed structures and explicit fields are one way to make the threshold legible. Not a solution. Just a way to leave a trail.
Asgeir Albretsen is the founder of Harbor.