Partly. When people first hear the term, they usually ask whether a world model for knowledge work is just a knowledge graph, and the structure is the same: things, the kind of thing each one is and how they relate. What a knowledge graph on its own does not say is what an agent needs before it acts. When was this fact true? Where did it come from? Is it accepted, or is it an agent’s guess? A world model for knowledge work is a knowledge graph with those answers attached to every fact, covering the world outside the company as well as the records inside it, and kept current by the work itself. If you already have a knowledge graph, you have the foundation.

A knowledge graph gives an agent things instead of strings

Google introduced its Knowledge Graph in 2012 as a model of “real-world entities and their relationships to one another: things, not strings”. At launch it held more than 500 million objects and 3.5 billion facts about them. That idea is the part we keep, and it does most of the early work.

At a deep-tech company selling into hospitals through integrators, the partner list was a spreadsheet of about 2,700 rows. Once the territory was modelled as a graph from government sources, hospital lists and integrator registries and linked to the CRM, it collapsed to 64 ranked accounts. The ranking came from a relationship, which integrator supplies which hospital, that no single source held. That is a knowledge graph doing its job. The ontology names the kinds of thing and the relationships allowed between them, and the graph holds the particular hospitals and people.

Most of that graph came from registries and lists outside the company. A graph built only from the company’s own systems holds the company’s own view of its market. The value was in joining that view to the world it describes.

Facts need time, because the business changes under them

A knowledge graph can say that a person works at a company. An agent also needs to know since when, whether it is still true and what it replaced. At a planning-software company, a new head of sales changed the deal stages without telling the CFO. The CRM just showed the new stages, so an agent working from it that week would have treated them as if they had always been there.

So in a world model for knowledge work, a new fact supersedes the old one instead of overwriting it. Each fact records when it was true and when it was recorded. When a person changes role, the new role becomes current and the old one stays as history. “What is true now” and “what did we believe last week” are both answerable, and they have different answers.

Knowledge graph

Plain edges

Every edge looks the same.

Person → works at → Company

Company → uses → Planner

World model for knowledge work

Edges with time, source and status

Person → works at → Company · from Sep, current · old role kept · source CRM · accepted

Company → uses → Planner · source GitHub commits · inferred, awaits promotion

The same two facts. The knowledge graph stores both as plain edges. The world model for knowledge work records when each was true, where it came from and whether it is accepted or inferred.

Every fact needs a source the agent can weigh

A company’s address from a government registry and the same address from a rep’s note are not the same fact, even when they match. The first is a record. The second is a guess that happens to agree. If the graph stores only the address, the agent cannot tell which one to trust when the two disagree.

In a world model for knowledge work every promoted fact carries its source, and the evidence stays underneath the state derived from it. An agent can rank a registry above a rep’s note. A person checking the agent’s work can follow any fact back to where it came from.

An inference stays an inference until someone accepts it

The planning-software company wanted to find companies still running an end-of-life open-source planner. GitHub knows which people committed to which repositories. The graph knows which company each person works at. Chain the two and you get “this company uses this software”, a fact no source states. Several people from one company committing was a reasonably safe bet. Stars alone were not, because people star repositories to bookmark them.

That conclusion is useful, and it is still a conclusion. A knowledge graph that writes it in as a plain edge makes it look like a record. We keep interpretation separate from accepted truth. The inference attaches to the company with its evidence, and it becomes an accepted fact only when a person or an agreed rule promotes it. Agents will write most of what enters the model, and without this line it fills with guesses that look like records.

The work keeps the model current

A graph that someone builds once describes the company on the day it was built. A world model for knowledge work is kept current by the agents that work from it. At the deep-tech company a top-50 list refreshes every morning at 7am. Each run leaves a trace on the graph, so the next run starts where the last one stopped, and as accounts move stage new ones fill in. Those traces are the working layer that some now call a context graph. The meeting-prep agent reads the same graph at 9am, so a correction made for one job is there for the next.

A knowledge graph says what exists. A world model for knowledge work also says what is true now, how we know and who agreed.