> Site index: [llms.txt](https://ortelian.com/llms.txt) with all pages and descriptions.

# Context graph vs world model

Source: https://ortelian.com/notes/context-graph-vs-world-model/
Markdown: https://ortelian.com/notes/context-graph-vs-world-model.md
Author: Laurens Nys
Published: 2026-10-14

A world model is the maintained company map over time. A context graph is the task-scoped slice agents get for one job.

A world model for knowledge work is the maintained map of your company and the world around it over time. A context graph is the slice of that map one task needs, plus the decisions already made about it. Vendors often treat the two as the same thing. They are not. One is the company model you keep. The other is what an agent is handed for a single job.

The plain definitions of a context graph are already owned by others. [Neo4j](https://neo4j.com/blog/agentic-ai/what-is-a-context-graph/) describes a connected representation of knowledge, state, history and decision records for a task. [IBM](https://www.ibm.com/think/topics/context-graph) uses it for structuring information inside a model context window. [Foundation Capital](https://foundationcapital.com/context-graphs-ais-new-layer-of-truth/) focuses on decision traces across entities and time. Our angle is narrower and more practical: the world model is what you maintain; the context graph is what each agent gets from it. For the full definitions, see [what is a world model for knowledge work](https://ortelian.com/notes/what-is-a-world-model-for-knowledge-work/) and [what is a context graph for AI agents](https://ortelian.com/notes/what-is-a-context-graph/).

## What is a world model?

A [world model for knowledge work](https://ortelian.com/notes/what-is-a-world-model-for-knowledge-work/) is a sourced, current map of the company and the world it operates in. It has three layers: the data that keeps it current, a graph of the things that matter and how they relate, and the knowledge attached to those things. Every fact keeps its source. Facts that can change also record when they were true, so an old role or price stays as history when a new one supersedes it.

It is maintained continuously. Systems of record and outside sources feed it. Agents and people add findings. Accepted facts stay separate from guesses until a person or an agreed rule promotes them. The whole company can share it. The next job starts from what the last one learned.

## What is a context graph?

A [context graph](https://ortelian.com/notes/what-is-a-context-graph/) is the slice one piece of work needs: the things involved, their current state and the decision traces behind them, each with a source. It is assembled for a task, not for the whole company. When an agent prepares a renewal, the context graph holds that customer, that contract, the current policy, the incidents that matter and the earlier exception that still counts as precedent. It does not hold every other customer in the territory.

Decision traces are what make a context graph more than a snapshot. A trace records what was decided, about which things, on what evidence, under which rule, by whom and when. Those traces only work if the things they point at are stable identities in a world model. Without that, a decision attaches to a string, and the same customer ends up with three partial histories.

## Context graph vs world model at a glance

|  | World model | Context graph |
| --- | --- | --- |
| What it is | The maintained map of the company and its world | The task-scoped slice drawn from that map |
| Scope | The company and the world it works in | One piece of work |
| How long it lives | Continuously maintained | Assembled per task; its decision traces are kept |
| What it answers | What is true now, and how do we know? | What does this job need, and what was decided before? |
| Who uses it | Agents and people across jobs | The agent (and reviewer) for this job |
| Built from | Systems, outside sources, findings and promotions | Identities, state and traces from the world model |

The world model is the map. The context graph is the route an agent is given for one trip. More context is not better: past a point, [more context makes agents worse](https://ortelian.com/notes/more-context-makes-agents-worse/). The world model is what lets you cut the right slice.

## How they relate to a knowledge graph and an ontology

These four terms sit in layers, and mixing them up is why buyers get sold the wrong thing.

| Term | Role |
| --- | --- |
| Ontology | The kinds of things that can exist, what they mean, how they may relate and what can be done with them |
| Knowledge graph | The particular companies, people, places and events, connected according to the ontology |
| World model | The graph plus the data that keeps it current, time and a source on every fact, and a line between accepted facts and guesses |
| Context graph | The slice one task needs, plus current state and decision traces |

A knowledge graph is roughly the middle layer of a world model. The longer compare is in [world model vs knowledge graph](https://ortelian.com/notes/is-a-world-model-for-knowledge-work-a-knowledge-graph/). The [ontology](https://ortelian.com/notes/what-is-an-ontology/) is the legend. [Building one](https://ortelian.com/notes/how-to-build-an-ontology-for-ai-agents/) starts from one piece of work, not from a company-wide catalogue.

## A sales example that needs both

At a deep-tech company selling into hospitals through integrators, a partner spreadsheet of about 2,700 companies collapsed to 64 ranked accounts once the territory was modelled from government sources and linked to the CRM. That work is world model work: one identity per hospital and integrator, sources joined, ranking on current facts. The public write-up is [one week at one company](https://ortelian.com/notes/one-week-at-one-company/).

Once that model exists, a meeting-prep agent does not need the whole territory. It needs the account for today's call, the people attending, the open deal, recent calls and the payment terms finance already agreed. That slice is the context graph for that meeting. The daily account list is another slice: net-new companies only, with a 30 to 60 day hold so the same name does not reappear. Each run leaves a trace, so the next run starts where the last one stopped.

Without the world model, both agents would search the CRM, inbox and Slack separately and guess which similarly named companies are the same customer. With only a context graph and no maintained model underneath, decision traces attach to strings and go stale as roles and policies change.

## What each one is for

Use a world model when agents (or people) need a shared, current picture across jobs: territory, customers, products, policies, facilities, suppliers. Use it when the same fact must mean the same thing in sales, operations and finance, and when yesterday's answer must not silently become today's.

Use a context graph when you are handing an agent one job. Cut the view to what that job needs. Include the decision traces that still apply. Leave the rest out. That is how you avoid stuffing the prompt with the whole company and calling it context.

You almost always need both if agents will act, not only answer. A context graph without a world model is a notebook of decisions about things that are not stable. A world model without a context graph dumps the whole map on every agent and [degrades the result](https://ortelian.com/notes/more-context-makes-agents-worse/).

## What goes wrong when the two are mixed up

Teams that buy "a context graph" and expect a company model often get a task log with no identity resolution, no time on facts and no promotion gate for agent findings. Teams that buy "a knowledge graph" and expect agents to work often get structure without current state or decision traces. Both miss the split: maintain the map; hand each job its slice.

Another failure is treating every agent conclusion as a fact. Agents will write most of the traces, because they sit in the path of the work. If every conclusion becomes precedent, the graph fills with guesses that look like history. Keep findings apart from accepted facts until a person or an agreed rule promotes them. That rule lives in the world model, and the context graph inherits it.

## How we build them at Ortelian

Ortelian is a platform plus a forward deployed team. We build the world model from the work, not from a blank company diagram. The audit takes about two weeks on site. We agree the standard on real examples, redesign the work, then build in two-week cycles. For each job that goes live, the agent gets a context graph cut to that job from the same world model. After launch, your team runs it or we keep running it. Either way we host and operate the platform in the EU.

For the deployment sequence, see [how to deploy AI agents in an enterprise](https://ortelian.com/notes/how-to-deploy-ai-agents-in-an-enterprise/). For why agents fail when they only get tools and no map, see [why AI agents give confident wrong answers](https://ortelian.com/notes/why-ai-agents-give-confident-wrong-answers/) and [why AI agents fail in production](https://ortelian.com/notes/why-ai-agents-fail-in-production/).

## Questions people ask

### Is a context graph the same as a world model?

No. A world model is the maintained map of the company over time. A context graph is the task-scoped slice drawn from it for one job, plus the decisions behind that job.

### Do I need a world model before a context graph?

You need what a world model provides: one identity per thing, facts with time and source, and a line between findings and accepted facts. Without that, traces attach to strings and go stale.

### Is a context graph just agent memory?

They overlap. Agent memory is what one agent keeps between runs. A context graph is shared, tied to company identities and keeps decisions with their sources. An agent's memory can feed it, but only as findings until they are promoted.

### Where do Neo4j, IBM and Foundation Capital fit?

They describe useful parts of the same stack: connected task context, structured context windows and decision traces. Our split is about ownership and scope: you maintain a world model; you hand each agent a context graph.

### Can chat tools replace either one?

Chat tools are strong at drafting and answering. They do not maintain one identity per customer across systems, keep time on every fact or cut a task-scoped slice with decision traces. Ortelian plugs into [ChatGPT Enterprise](https://ortelian.com/ortelian-vs-chatgpt-enterprise/) and [Microsoft Copilot](https://ortelian.com/ortelian-vs-microsoft-copilot/) inside an engagement. The world model and context graphs stay in the platform underneath.