What is unwritten knowledge?

Unwritten knowledge is everything an organization knows but never recorded: the skill, judgment, relationships, and situated context that live in people rather than in documents. It is most of what a company actually runs on, and almost none of it is in the files.

It goes by many names, each catching a different angle — tacit knowledge (the know-how you can't put into words), tribal knowledge (what the team shares but never wrote), institutional memory (the why behind decisions). The slice that matters most in the AI era has a sharper name: dark context.

Dark context is what an organization knows but never wrote down — the situated context a general AI model cannot see. A general language model holds all the general knowledge and none of yours; that gap is your dark context.

That is why unwritten knowledge is suddenly worth naming precisely. For decades it was a management problem — onboarding, succession, "sit next to Maria for a month." Now it is the exact thing standing between a fluent AI and one that actually knows your business. Read the Dark Context manifesto →

People treat these words as rough synonyms and then argue about which is correct. That is the wrong shape of question. They cover almost the same ground; what differs is which question each one was built to answer. Pick the term by picking the question.

Tacit knowledge — why can't it be told?

The oldest of the five, and the only one that is really about the limits of language. It names knowledge that resists being said at all, not knowledge that merely went unsaid. Scale: one person. Field: philosophy, then management.

Tribal knowledge — who holds it?

Shifts the question from the mind to the group. The knowledge could be written; it isn't, because the team passes it along by proximity instead. Scale: a team. Field: organizational practice. A thing can be tribal without being tacit — perfectly sayable, just never said to anyone outside the room.

Institutional memory — what does it hold?

The only one of the five that is explicitly about time. It names the accumulated reasoning behind decisions — not how to do the work, but why the work is set up this way. Scale: the organization, across years.

Dark context — can a machine see it?

The newest, and the only one that asks a machine-readability question rather than a human one. The other four all predate the thing that made this urgent: none of them was built to answer whether a model can use what you know. That is why a new name was needed rather than a fifth borrowing of an old one.

Unwritten knowledge — the umbrella

The term that declines to commit to any one dimension, which is exactly what makes it useful as a starting point and useless as a diagnosis.

So the disagreement about vocabulary is usually a disagreement about which problem is in the room. If you are arguing about whether something is tacit or tribal, you are really arguing about whether the fix is elicitation or documentation — and that is worth arguing about.

Nobody arrives here having decided they have a tacit-knowledge problem. They arrive with something going wrong. Find yours.

"Someone with twenty years here leaves in six weeks."

The deadline case. Everything else on this page becomes urgent when a specific person hands in notice, because the window to ask them anything is now finite.

"Our AI writes beautifully and knows nothing about us."

The fluency trap. A general model holds every business in general and yours not at all, and fluency makes the gap harder to see rather than easier.

"We pointed a chatbot at our wiki and it is confidently wrong."

You reached for retrieval and it surfaced everything anyone wrote — which is not the same set as everything anyone knows.

"We have the data. The AI still can't use it."

Two different blind spots wearing one word. One was collected and ignored; the other was never collected. They need opposite fixes, and observability closes neither.

"New hires take nine months and nobody can say why."

What they are slowly absorbing is real, load-bearing, and written down nowhere. The onboarding document is not the thing they are actually learning.

"We keep re-litigating a decision nobody remembers making."

The rule survived; the reasoning behind it did not. So every year someone proposes removing it and nobody can say what breaks.

"Our best person can do it but can't teach it."

The hardest case, and the oldest. It was never withheld and never written, because it was never fully sayable in the first place.

Getting unwritten knowledge into words is the first move. Keeping it true is the second, because captured knowledge decays. Context has a lifecycle: green (current, findable, you can ask it), gray (was true, now drifting), and dark (gone). See the seven context levels and how context decays.

What is unwritten knowledge?

Everything an organization knows but never recorded — skill, judgment, relationships, and context that live in people, not documents. Tacit knowledge, tribal knowledge, and institutional memory are all names for parts of it.

Is unwritten knowledge the same as tacit knowledge?

Tacit knowledge is the classic term for the individual, hard-to-articulate kind — Polanyi's "we know more than we can tell." Unwritten knowledge is the broader umbrella, and dark context is the organizational slice of it that a general AI model can't see.

What is the difference between tacit knowledge and tribal knowledge?

Tacit knowledge asks why something cannot be told: it resists words, so the person holding it often cannot teach it directly. Tribal knowledge asks who holds it: a team passes it along informally, and it could be written down — it simply never was. Something can be tribal without being tacit, and the difference decides the fix. Tacit needs elicitation; tribal needs documentation.

Is dark context just a new name for tacit knowledge?

No, and the difference is the question each one asks. Tacit knowledge asks a human question — can this be put into words at all? Dark context asks a machine one — can a model use what your organization knows? All four older terms predate the thing that made this urgent, which is why the newer one exists rather than being another borrowing.

Why does unwritten knowledge matter for AI?

A language model learns from the written record, so it holds the world's general knowledge and none of your organization's unwritten know-how. Retrieval doesn't help: you can't index what no one wrote down. Someone has to put it into words first.

How do you capture unwritten knowledge without losing the person who holds it?

By treating the holder as the keeper, not a bottleneck to route around: surface the context they carry in their own words, keep it current, and credit them for it. That is the work Dark Context is about.