POV | What We Believe, On the Record | Mactores

What Is a Context Graph vs a Knowledge Graph and Ontology?

Written by Gautham B | Nov 16, 2026, 12:20:58 PM

Key Takeaways

  • A context graph records how decisions were made, including who or which agent decided, which sources and rules were used, and which lots, customers or orders the decision touched.
  • A knowledge graph holds what is true now, while a context graph holds what was decided and on what basis. Both are made of nodes and relationships, but they answer different questions.
  • The ontology defines the terms both graphs use, so a context graph sits next to it rather than replacing it, and traces from different plants, teams and agents stay comparable.
  • The coatings maker in this guide shows why source versions matter. A lot released against one revision of a customer's specification stopped being a precedent once that specification tightened.
  • Traces have to be written when the decision happens. Rebuilding them later from email and logs rarely recovers which version of each source someone relied on.

If you've only recently heard the term context graph, you're not late. It moved from venture capital essays into enterprise AI roadmaps over the past year, and vendors still use it for different things. The most useful meaning is the narrow one: a context graph records how decisions were made and which sources were used, so people and AI agents can find precedent and auditors can see why something happened. That places it next to the ontology and the knowledge graph, not in place of either.

Our running example is a specialty coatings manufacturer, made up but typical, with two plants that make primers for auto parts makers. Early in 2026 it put an AI agent on its quality release queue. When a lot fails an internal test, the agent drafts a disposition and a quality manager approves or overrides it. One override in September, and what happened to the next lot five weeks later, shows what a context graph is for better than a definition can.

What Is a Context Graph?

At the coatings maker, the context graph is the record that says lot 26-0917 shipped under a deviation, who approved it, which test results and which revision of the customer's specification the decision relied on, and what the approval covered.

In general terms, a context graph is a graph of decisions. Each decision is linked to the people or agents who made it, the sources and rules they used, and the business entities it affected.

The lot itself was ordinary. PR-410 primer from plant 2 measured 4,850 centipoise on the viscosity test, against an internal limit of 4,500, and a retest read 4,820. The agent drafted a hold for rework. The quality manager overrode it because the auto parts customer's specification, revision C, allowed up to 5,000, and she approved release under deviation DV-0417 for that lot and that customer only.

Part of the trace

What DV-0417 recorded at the coatings maker

The decision

Release lot 26-0917 under deviation DV-0417, September 9

Who decided

The agent drafted a hold, and the plant 2 quality manager overrode it

Sources used

Viscosity test at 4,850 cP, retest at 4,820 cP, the customer's specification revision C

Rule applied

Deviation procedure QP-12, version 4

Entities touched

Lot 26-0917, PR-410 primer, plant 2, the auto parts customer

Conditions

This lot and this customer only, with the customer notified the same day

The reason given

"Within the customer's 5,000 cP limit under specification revision C"

The phrase took on this meaning in a December 2025 essay by Jaya Gupta and Ashu Garg of Foundation Capital, which describes a context graph as "a living record of decision traces stitched across entities and time" so that precedent becomes searchable. Some vendors use the same words for the subgraph an AI system loads into a model's prompt, or for a knowledge graph scoped to one company's data. Those things are real, but they already have names, and neither one records why a decision went the way it did.


Context Graph vs Knowledge Graph

A knowledge graph records what is true about your business, and a context graph records what was decided and on what basis. Both are nodes and relationships, which is why the terms get blurred. The difference shows up as soon as something changes.

The coatings maker's knowledge graph knows that lot 26-0917 is PR-410 primer made at plant 2 and sold to the auto parts customer, whose current specification is revision D. On October 1 the customer tightened its viscosity limit to 4,600, and the knowledge graph updated, as it should. It had no reason to keep revision C. The context graph did, because DV-0417 was approved against revision C, and that was the part of the decision that stopped being true.

Question at the coatings maker

The knowledge graph

The context graph

"What's this customer's viscosity limit?"

4,600 cP, under revision D

Not its job

"Have we shipped this customer a lot above our own limit?"

Only if someone stored it as a fact

Yes, under DV-0417 in September

"What did that release rely on?"

No answer

Two test results and revision C

"Who overrode the agent, and why?"

No answer

The plant 2 quality manager, within the customer's limit at the time

"Does that decision still apply?"

Can't say on its own

Not since October 1, when revision D replaced revision C

If a vendor describes its context graph as a knowledge graph scoped to your company, it's worth asking where the decisions live. A separate audit log can keep the record, but precedent is hard to search when the log doesn't point at the lots, customers and specification revisions involved.

Where the Ontology Fits

The ontology defines the terms both graphs use, so a context graph sits next to it, not in place of it. At the coatings maker, the ontology says what a lot, a specification revision and a deviation are, how they relate and which rules apply, such as a deviation covering one lot for one customer and a newer specification revision replacing the old one.

That agreed model of the business is what ontology in AI work produces. The knowledge graph fills those terms with facts, and the context graph records decisions in the same terms, along with the ontology version that was in force.

Without that shared vocabulary, traces pile up but don't line up. Before the ontology, plant 1 logged out-of-spec releases as concessions and plant 2 logged them as deviations, with different reason codes, so a search for precedent at one plant missed the other plant's history. Mapping both to one deviation class made 11 earlier releases at plant 1 visible to the agent.

The two also feed each other. When the same exception keeps being approved for the same reason, the rule is probably wrong, and its owner can change it in the ontology instead of approving it case by case. When a rule does change, the context graph shows which past decisions would have gone differently.

None of this needs a new standard. The W3C's PROV-O ontology, a Recommendation since 2013, already models provenance as entities, activities and agents, and a decision fits it well: an activity that used entities such as test results and a specification, was associated with an agent, human or AI, and produced an outcome. PROV-O is itself an ontology, which is one more reason the two belong together.


How Agents Use a Context Graph

Agents use a context graph in two directions. Before they act, they look for precedent and check whether its basis still holds. After a decision, they write a new trace, or the workflow around them writes it.

On October 14, lot 26-1102 of the same primer read 4,780 centipoise. Without the context graph, the agent would have drafted the same hold it drafted in September, and a quality manager who remembered DV-0417 could easily have released the lot again on the old reasoning.

With it, the agent found DV-0417 and read its basis. It saw that revision C had been replaced by revision D, with a limit of 4,600, and drafted a hold anyway, this time with both revisions side by side and a note that the September precedent no longer applied. The quality manager approved the hold in a few minutes, and the lot went to rework.

The trace for that hold was written as part of the approval, not pieced together afterward. A trace rebuilt a week later from email, logs and memory rarely captures which revision of each source someone looked at, and here the revision was the whole point. Traces can live in a graph database or in the same governed tables as the data they describe, which is how we approach a data lakehouse for agentic AI, as long as each one points at the entities and versions it used.

 

A context graph needs the same governance as the data it points at. Traces can include customer names, contract terms and sometimes personal data, so they inherit access rules, carry a retention period and need a way to retire precedent that no longer applies. The context graph is one part of a context layer for enterprise AI, alongside the ontology, governed definitions and the systems of record the agent reads.

In the context ontology services we deliver, each agent answer already records the definitions, sources, rules and ontology version behind it, which covers most of what a context graph needs. The rest is linking those records to the business entities they touched and making them searchable before the next decision.

Context Graph Examples

Any decision that recurs, involves judgment and gets questioned later is a good candidate. The pattern is the same in each industry: the decision, the sources and rules behind it, and the precedent question an agent would need answered before it acts on the next case.

Industry

The decision

What the trace keeps

What an agent would ask next time

Manufacturing

Releasing a lot under deviation

Test results, specification revision, approver, scope

"Has this customer accepted a lot like this, under the spec it has today?"

Banking

Closing an anti-money-laundering alert

Parties, ownership links, rule version, the investigator's rationale

"Why did we close the last alerts on this group of accounts?"

Health insurance

Approving a prior authorization on appeal

Clinical records read, policy version, reviewer

"Which version of the policy did the reviewer apply?"

Media

Clearing a title to stream in a territory

License terms, territory, dates, rights-holder approval

"Under which license did we stream this title in Canada last spring?"

Software

Escalating a support case to engineering

Incident history, customer tier, the on-call engineer's call

"What did we do the last time this customer hit this error?"

In regulated industries, someone usually asks for this record already. For banks, our context ontology for financial services work stores every conclusion with the data it read, the rules it applied and the ontology release in force.

Rights holders can audit a streaming decision long after the title has aired, which is why our context ontology for telecom and media design keeps a trace per answer for regulators and rights holders.

Most teams should start with one decision type rather than a company-wide graph. At the coatings maker it was out-of-spec lot release, because it came up every few weeks, an agent was already drafting it and customers' quality teams asked about it. The ontology for lots, specifications and deviations came first, the agent wrote traces from its first day on the queue, and the precedent check went live in October once enough traces had built up.