All notes

World models · · 4 min read

What is a context graph?

The short answer

A context graph is the record of what a company decided about its work, why, and who approved it, tied to the customers, orders, policies and people those decisions are about. It sits inside the company world model. The world model says what is true now. The context graph says how the company got there and what counts as precedent.

What a system of record leaves out

A CRM, an ERP or a ticketing tool keeps the current state. The order is confirmed. The discount is 15%. The supplier is approved. What these systems rarely keep is the reasoning: which rule applied, which exception was made, what evidence it rested on and who said yes. That part lives in email threads, chat messages, call notes and people's memory.

For a person who has been at the company for years, that is fine. They remember. For an AI agent, it is the gap that makes it guess. An agent that sees a 15% discount where policy says 10% cannot tell whether that is an error, a one off or a standing agreement. A context graph is what lets it tell the difference.

What goes into a context graph

Each entry is one decision, attached to the things it is about.

PartWhat it recordsExample from an order desk
DecisionWhat was decidedAccept a delivery date two weeks earlier than the standard lead time
AboutThe things it concerns, each as one identityThis customer, this order, the lead time rule
EvidenceWhat it rested on, with sourcesThe planning sheet showed free capacity that week
RuleThe rule applied, or the exception to itStandard lead time of six weeks, exception granted
Approved byA person, or an agreed ruleThe production planner
WhenThe date, and which version of the rule was in forceThe order date, the lead time rule current then

The "About" row is where most attempts fail. If the customer is a text string, the same customer spelled three ways becomes three histories. That is why the context graph is part of the world model rather than a separate log: the world model holds one identity per customer, person, product and policy, and every decision attaches to those identities.

How a context graph fits in the world model

A world model for knowledge work is one current, sourced picture of a company and the world it works in. On the Ortelian platform it has layers: the ontology defined with the client, the state of companies, people and deals, what people know and agents learn, and how the work turned out. The context graph is the part that records decisions and outcomes over time.

When an agent picks up one job, it is handed the slice of the world model that job needs, including the decisions that still apply. The context graph vs world model note covers that split: the world model is what you maintain; each job gets its slice. For how the four terms sit together, see context graph vs knowledge graph vs ontology.

Why agents need it

Three things go wrong without it.

  1. Agents repeat old exceptions as if they were rules. A discount approved once for a reason that no longer holds becomes the new default.
  2. Agents treat their own guesses as history. If every conclusion an agent writes is stored as fact, the record fills with guesses that look like precedent.
  3. Nobody can check the agent. Without a record of why, a reviewer has to reconstruct the reasoning from scratch every time.

We keep an agent's findings apart from accepted facts. A finding attaches to the thing it is about, with its evidence. It becomes accepted only when a person or an agreed rule promotes it.

How Ortelian builds one

We build it as part of the world model, for one piece of work at a time, from the systems the company already runs. The steps for that are in how to build a context graph from the systems you already have.

On the platform, every fact keeps its source, and every run, action and approval is recorded, with the sources behind it. Actions that leave the company, such as sending an email or an invoice, wait for a person's approval by default, and approval points are set before a workflow goes live. What one agent learns, the next can use. The platform is hosted in the EU.

Questions people ask

Is a context graph the same as a knowledge graph?

No. A knowledge graph holds what exists and how it connects. A context graph adds the decisions made about those things, with their evidence, rule, approver and date.

Is a context graph separate from the world model?

Not in how we build it. The context graph is the part of the world model that records decisions and outcomes. Keeping it inside the world model means every decision points at one stable identity.

Who writes the entries?

Agents write most of them, because they are in the path of the work. People approve the ones that need approval, and an agent's conclusion stays a finding until a person or an agreed rule accepts it.

Do you need to move data out of your ERP or CRM?

No. The systems of record stay where they are. The build guide covers which system holds which part.

View Markdown

We take on three companies per quarter.

The first two weeks start that work on one job.

Antwerp from across the Scheldt, with the Cathedral of Our Lady and the Boerentoren.