Last updated: 2026-10-10

U
Undergraduate level

Where Does an Organisation's Knowledge Live?

A database, a knowledge graph, and a large language model each answer a simple question — where is the knowledge, and how is it found again — in genuinely different ways. An organisation turns out to answer it in a fourth way, one that borrows a little from each of the other three without being reducible to any of them. Seeing the comparison clearly, including where it breaks down, is the starting point for the rest of this part of the cluster.

Four Architectures, Briefly FoundationalKnowledge that endures for decades — core principles

A database stores explicit records under a deliberately designed schema. Its knowledge is addressable (each record has a location), inspectable (you can look at it directly), updateable, attributable (a record has an author and a timestamp), and governed by explicit integrity rules. What a database does not do is decide, for itself, what its records mean, which ones matter right now, or how two conflicting records should be reconciled — those judgements sit outside the database, in whatever larger system uses it. A knowledge graph goes a step further by making relations between records explicit and navigable rather than leaving them implicit in a schema design, which is this site's own introduction to the idea develops from the ground up. A large language model stores knowledge in neither of these ways: what it has learned is distributed across its parameters and expressed through context-sensitive reconstruction at generation time, rather than kept as a set of independently addressable propositions — this site's page on understanding LLMs covers the next-token mechanism this depends on.

The Organisational Equivalent of a Knowledge Graph FoundationalKnowledge that endures for decades — core principles

This cluster's earlier page already argues that what collapses when an expert leaves is retrieval rather than storage — the knowledge persists somewhere, but the fast, reliable path to it is gone. Read against a knowledge graph's own logic, that earlier argument says something more specific: an organisation retains a document but can still lose the edges connecting it to the rest of the graph — the link between the document and the problem it addressed, the vocabulary needed to find it again, the person who knew it was relevant, the context explaining why one exception mattered, confidence in whether it's still authoritative. The organisational equivalent of a knowledge graph's nodes and edges, then, is not only documents and formal relations. It includes people who act as high-value nodes, trust relations between them, communities of practice, shared classifications, remembered cases, escalation routes, and plain habits of asking the right person first. When an expert leaves, an organisation typically loses not a repository of facts but an index, a router, and an inference mechanism all at once — which is why the loss is felt immediately in what the organisation can no longer quickly do, well before anyone discovers a specific fact has actually gone missing.

Dispositional Knowledge: the LLM Parallel FoundationalKnowledge that endures for decades — core principles

Much of what an organisation knows is dispositional rather than declarative — visible in what it can reliably do, notice, or reconstruct, not necessarily recorded anywhere a person could point to. That's a genuinely illuminating parallel with a large language model, whose learned knowledge is similarly expressed through behaviour rather than stored as an inspectable list. Like a model, an organisation can reproduce a familiar answer without anyone being able to identify its original source; combine fragments into a plausible but historically inaccurate account; behave differently depending on how a question is framed and by whom; preserve a pattern of action while losing all record of its provenance; and lean on an external system — a wiki, a shared drive, a consultant — to supplement what it can no longer reconstruct from memory alone. None of this makes an organisation a language model. It means both kinds of system can hold knowledge that only becomes visible in use, and that neither kind can be fully audited by inspecting a store of records, because for both of them a great deal of what they "know" isn't represented as records at all.

Strongest Objection: The Comparison Can Mislead as Much as It Illuminates FoundationalKnowledge that endures for decades — core principles

Lining organisations up next to databases, graphs, and models risks treating a genuinely different kind of thing as a fourth member of the same family, when it's better described as something that contains elements of all three while adding what none of them have at all: embodied people with motives, a material environment, incentives, and power relations that shape which knowledge gets attended to in the first place. A database doesn't care who's asking. An organisation's retrieval paths are shaped by exactly that — this cluster's page on organisational roles already makes the case that who asks, and who's trusted to answer, is not incidental to what gets found. The four-way comparison earns its keep only as a way of separating which kind of knowledge-architecture question is being asked about an organisation at a given moment — not as a claim that an organisation simply is a database, a graph, or a model with extra steps.

Provisional Conclusion FoundationalKnowledge that endures for decades — core principles

An organisation's knowledge lives in a hybrid architecture that a database, a knowledge graph, and a large language model each only partially resemble: explicitly recorded and directly addressable like a database in places, available through structured relations like a graph in places, and distributed across dispositions that only emerge through reconstruction, like a model, in still other places — with embodied people, practices, and power relations holding the whole arrangement together in a way none of the other three architectures needs. Treating any single one of these as the full picture of where an organisation's knowledge lives will miss most of it.

Questions for Further Thought

  • If an organisation's knowledge is genuinely hybrid, is there a reliable way to tell, for a specific piece of institutional knowledge, which of the three architectures it currently most resembles?
  • Does deliberately converting dispositional organisational knowledge into something more database-like (a written record) always help, or does the conversion sometimes lose exactly the context-sensitivity that made the dispositional version useful in the first place?
  • Could an organisation's reliance on a specific person as a high-value retrieval node be measured directly, the way a graph's centrality measures could identify a critical node before it's lost?