The Knowledge You Find by Talking
Most PKM tools assume you have the thought before you write it down. They're wrong about when knowledge appears.
In 2008, NASA ran a project with an unusual premise. Engineers who'd spent careers on the Space Shuttle program were retiring, and NASA had realized that the documentation — hundreds of thousands of pages of procedures, schematics, and incident reports — wasn't capturing what was actually leaving when they walked out the door. Not the technical specifications. The judgment. The decision rationale. The lessons that had changed how teams worked after close calls that never made it into formal records, because writing them down would have required admitting something nearly went wrong.
So NASA flew engineers to its centers in Houston, Florida, Alabama, and Mississippi, sat them down, and interviewed them. The resulting archive, called the Space Shuttle Tacit Knowledge Capture Project, runs to hundreds of hours of recorded conversation. What they found wasn't hidden information. In most cases, it was information that hadn't existed in transmissible form until someone asked.
This is what knowledge management researchers call externalization — the moment when tacit knowledge becomes explicit. Ikujiro Nonaka and Hirotaka Takeuchi named it in 1995 in The Knowledge-Creating Company, as one of four fundamental modes of knowledge conversion. You know something. Then someone asks you about it. And in the act of explaining, the knowledge crystallizes into a form that can actually be written down, stored, shared. Before the question, it was pattern recognition. After, it's a sentence.
Most tools are designed for the wrong moment
Almost every note-taking and PKM system is built on a single implicit model: you have a thought, then you capture it. Thinking happens first. Writing is recording.
But anybody who has worked through a difficult problem by explaining it to a colleague knows this isn't quite right. The explanation isn't downstream of the thinking; it's part of it. You find out what you know by trying to say it. The gaps appear when you reach for the next sentence and discover there's nothing there.
Cognitive Task Analysis was designed around exactly this problem. Developed as a methodology for eliciting expert knowledge, CTA exists because domain experts often can't explain what they know when asked directly — not because they're hiding anything, but because their knowledge is stored as pattern recognition, not as rules. There are now over a hundred CTA methods in use, and all of them share one structural assumption: getting expert knowledge out requires conversation, not a blank page.
The protégé effect, which Nestojko et al. documented in 2014, shows the same pattern from a different angle. When people expect to explain something to someone else, they learn it differently — more deeply, with more attention to where their understanding breaks down. The anticipation of articulation changes how you process. Even imaginary audiences work: Fiorella and Mayer found in 2013 that students who recorded explanations for a fictitious future student still got meaningful learning benefits.
Rubber duck debugging is the programmer's folk version of this. The duck doesn't say anything. It doesn't need to. You find the bug in the act of explaining it.
What AI chat makes possible, and what it doesn't do
The recent wave of capable AI assistants changes something here, at least in principle. For the first time, you can think out loud with something that asks follow-up questions. You can work through a decision by talking it through, explain what you know and discover what you don't, have the AI probe the edges of your reasoning. The externalization mechanism that NASA sent engineers to Houston for is now available at your desk.
I've found myself doing this without intending to. An hour in a conversation where I'm working through a product decision, and by the end I understand my own position more clearly than I did going in. Something crystallized. The act of explaining it made it explicit in a way it wasn't before.
But by default, what comes out of that conversation goes nowhere useful. It lives in a chat log, which is optimized for someone who was present, reviewing a sequence of exchanges. It's not retrievable by future-me searching for a decision I made, or queryable by an AI that needs context about my current thinking. The knowledge appeared — and then disappeared back into a format that's structurally inappropriate for retrieval.
This is different from the note-taking problem. It's not that I failed to write something down. It's that the tool which produced the knowledge and the tool which stores the knowledge are different things, with no connection between them.
The translation that doesn't happen
The gap that matters isn't between what you know and what you wrote down. It's between what you found out by talking and what ended up somewhere you can use later. Those are genuinely different problems, and most knowledge workflows treat only the first one.
When something crystallizes in a conversation — a decision made, a preference clarified, a person's situation understood differently — there's a brief window where it exists in a form that could become a record. A structured entry in a knowledge base. A preference updated. A note with actual context. After the conversation ends, that window closes and reconstruction gets harder.
The harder design problem isn't getting AI to remember things. It's making the path from externalization to durable knowledge short enough that it actually gets used. Not everything that comes out of a conversation is worth keeping. But some of it is the best knowledge you have — knowledge that didn't exist before you explained it. Losing it to a chat log is a strange kind of waste.
Asgeir Albretsen is the founder of Harbor.