The Knowledge You Can't Write Down
Notes capture what you thought. They rarely capture why — and almost never the judgment that decided it.
In 1966, Michael Polanyi published a slim book called The Tacit Dimension that tried to explain something every expert already knew but couldn't quite name. His key claim fit in a single sentence: "We can know more than we can tell."
He meant it precisely. Not that we know things we're reluctant to share, but that a large part of what we actually know is unknowable in propositional form. You can't tell someone how to ride a bike; you can describe it, but the description is useless until the body figures out what the description can't convey. You can't explain how you recognize a face. You can't articulate what made one candidate clearly better than another after an interview — you walked in with a rubric and walked out with a gut feeling that overrode it.
Polanyi called this tacit knowledge. The knowledge that doesn't transfer through documentation.
What notes actually capture
Most personal knowledge tools are built on an implicit assumption: that knowledge is the kind of thing you can write down. This isn't entirely wrong. Explicit knowledge is real and worth keeping: facts, decisions, preferences, meeting notes, the things someone told you that you'd otherwise forget. Structured systems for storing this are useful.
But it's a partial picture. What makes explicit knowledge usable — letting you apply a decision to a new situation, knowing when a preference overrides another, sensing which fact belongs here — is tacit. It lives in you, not in your notes. Notes are the tip of an iceberg that's mostly underwater.
You can write down that a vendor was difficult to work with. You can't write down the specific kind of difficult, the pattern you've learned to recognize, the early signal that now sets off a warning you don't have a name for. Those signals accumulate over years. They don't live in a document. They live in your judgment.
Why this matters for AI memory
AI memory systems face the same limitation, more acutely.
A structured knowledge base can store a lot: your preferences, your person records, your project notes, decisions with their rationale, tasks with context. That's genuinely more useful than a blank start. But the knowledge it can hold is bounded by the same ceiling. What you know how to put into words.
What AI can't observe is your tacit expertise. It doesn't know that you pause slightly before agreeing to aggressive timelines. It can't see that you always find one specific person worth calling before a contract decision. It can't learn that your first instinct about a creative direction is usually right, even when you talk yourself out of it. These patterns are real knowledge. They're not in the database.
David Autor at MIT has argued that Polanyi's paradox explains which jobs proved most resistant to automation: work involving complex perception, judgment, and situational reading. You can automate what can be specified. You can't automate the knowledge that resists specification. The same ceiling applies to your personal tools.
What could change this
There are two possible ways through. Neither is clean.
The first is inference. If an AI observes enough of your behavior — your choices, your hesitations, what you approved and what you didn't — it can build a model of your tacit preferences that you never explicitly stated. This is how recommendation systems work. It's also how surveillance capitalism works, which is why this idea in a personal tool requires a level of trust that very few systems have earned. You'd need to be certain about what's observed, where it goes, and who controls it.
The second is better capture at the moment of decision. Not notes about what happened, but notes about why — the reasoning, the rejected alternatives, the feeling that tipped the balance. Most people don't do this because it's slow and feels performative. But a note written at the moment of judgment, before the tacit knowledge recedes into background, captures more than a note written later. Capturing rationale is harder than capturing outcomes, and far more valuable.
Neither fully solves the problem. The tacit dimension doesn't compress neatly. Polanyi spent his career arguing it wasn't just a practical limitation. It was structural. Some things can only be transmitted through practice, observation, and apprenticeship. A note is not an apprenticeship.
The honest version of what a knowledge base can do is narrower than the marketing tends to imply. It can preserve your explicit knowledge so you stop losing it. It can give AI tools enough context to be genuinely useful rather than generically plausible. That's worth building. But it's the part above the waterline.
The part that makes you actually good at things is somewhere else.
Asgeir Albretsen is the founder of Harbor.