"We know more than we can tell." Michael Polanyi's six-word claim, made in 1966, is the single most consequential idea in organizational learning. It is also the single most-ignored idea in operational knowledge-management practice.
A bicycle. Most people who can ride one cannot, in any useful sense, explain how. Asked to write down "how to ride a bicycle," they will produce something - pedals, balance, lean into the turn - that is recognizably about bicycling but would not, when handed to a novice, suffice. The novice will fall. The novice has the document. The document is not the knowledge.
This is the entire problem of tacit knowledge in twenty seconds. The bicyclist knows how to ride; their body knows it, their inner ear knows it, their reflexes know it. None of that knowledge is propositional. It does not reduce to bullet points. It is real and it is unwriteable. Polanyi's term for the thing the bicyclist has and the document does not is tacit knowledge, and the gap between the two is the explanatory engine of nearly every failure of institutional knowledge management.
Michael Polanyi published The Tacit Dimension in 1966. The book is short and difficult, and the central thesis is contained in its most-quoted line: we know more than we can tell. Polanyi was a chemist before he became a philosopher, and his argument was empirical, not abstract. He pointed out that scientific practice itself depends on tacit components that resist formalization - a chemist's sense for which experiments are worth running, a doctor's clinical intuition, a craftsman's feel for the material. These judgments are real, they are reliable, and they are not, in any complete sense, articulable. They are passed from teacher to apprentice through a process Polanyi described as indwelling - a long, embodied participation in a practice - rather than through transmission of facts.
The implication for organizations is severe. If a meaningful share of what an institution knows is held tacitly - in the bodies, habits, instincts, and unstated judgments of its experienced practitioners - then the standard institutional response to knowledge loss (write it down, put it in the wiki, train the new person on the document) cannot, in principle, capture the part that matters. The institutional knowledge problem is not, primarily, a documentation problem. It is a transmission problem in a medium that does not admit of writing.
Polanyi's distinction is binary - explicit versus tacit - but the more useful taxonomy is finer. Massingham (2018), drawing on Spender and others, divides knowledge along two axes: explicit/implicit and individual/social. The four resulting quadrants are the standard frame in the contemporary knowledge-management literature:
What an individual knows and could write down if asked. The contents of their working memory. Survives most departures, in principle, because it can be elicited.
What the team has already written down - wikis, runbooks, decision records. The "explicit corporate memory" most KM tools target. This is the easy part.
The expert's muscle memory, the operator's reflex, the writer's ear. Real, fast, reliable - and almost entirely unconscious. The bicyclist's knowledge sits here.
"How we do things here" - the unwritten rules a culture participates in but no member can fully articulate. The hardest to capture and the easiest to lose in reorgs and acquisitions.
Quadrants 1 and 2 are the explicit half. Quadrants 3 and 4 are tacit. The first half is what institutional knowledge-management tools have been built to address since the 1990s. The second half is where organizational amnesia, in our longitudinal data, disproportionately concentrates. When a senior engineer leaves, the institution loses some Q1 (which was retrievable but not retrieved), some Q2 (rarely affected), most of their Q3 (lost entirely), and a real fraction of Q4 (because their participation in the unwritten rules ends).
The failure is not lack of effort. The failure is structural. Three reasons, from the literature and from our own departure-interview field work:
The expert cannot access their own tacit knowledge on demand. Asked "what do you know about handling X?", an experienced practitioner produces an outline of the explicit (Q1) component. The implicit (Q3) component just isn't available to introspection - it's the ground the introspection itself stands on. The bicyclist asked to introspect about balance produces inadequate metaphors; the surgeon asked to introspect about technique produces a textbook. The thing that makes them an expert is precisely the part they cannot describe.
Documentation captures conclusions, not process. A runbook says "if X, do Y." It does not, in general, say "you will know X is happening because of these three signals you would not have noticed before three years of pattern-matching, and you will choose Y over the seven nearby Y-prime options because of considerations the room would have to be in to articulate." The reasoning is the load-bearing part. Documentation routinely strips it out, because the experts writing the documentation are not aware of the reasoning.
The reader cannot reconstitute tacit knowledge from explicit text. Even a perfect transcription of an expert's tacit knowledge - if such a thing were possible - would not, when read by a novice, produce expertise. Tacit knowledge is acquired through what Polanyi called indwelling: extended practice within the form. The novice who reads the bicyclist's perfect transcript still falls. They need to ride.
The literature converges on three mechanisms, each ancient, none cheap:
Each of these mechanisms takes time. None of them are software. All of them are, in our field data, observable in their absence - the institutions whose tacit knowledge is most at risk are precisely the ones that have allowed apprenticeship, community-of-practice, and situated-reflection rituals to atrophy. The remedy is not to deploy a tool. The remedy is to reinstate the ritual.
The current wave of AI-memory work - RAG, vector stores, agent memory - sits, in Polanyi's terms, almost entirely in Quadrants 1 and 2. AI memory systems index and retrieve explicit knowledge. They do this well, sometimes spectacularly. They do not, on our current evidence, capture tacit knowledge - and the more confident the model becomes about answering questions whose answers required tacit knowledge to formulate, the more dangerous the failure mode. A RAG system that retrieves the right document and wraps it in fluent prose can hide the absence of the underlying tacit understanding from the user, who has no way to verify whether the answer was reasoned-from-experience or pattern-matched-from-text.
This does not mean AI memory is useless for organizational knowledge - it means the explicit/tacit distinction needs to be designed for, not papered over. Systems that flag when a question's reliable answer would have required tacit knowledge - and route it to a human practitioner instead of fabricating a confident response - are likely to be the more durable architectures. The literature on tacit knowledge, sixty years old this year, is unusually relevant to a problem the AI community thinks of as new.
Working primer; not peer-reviewed. Feedback at pb@reattend.ai.
The institution that forgets its tacit knowledge does not forget anything in particular. It forgets how to recognize what it has forgotten.