Context graphs vs knowledge graphs: what actually differs, and the ontology underneath

Two vendors shipped products in the last two weeks of October, and four more published around the same idea. Neo4j launched a Context Engine on October 20. Kore.ai published a context-graph explainer the same month. DataHub defined the term. AllegroGraph shipped a product around it. Atlassian exposed its Teamwork Graph. When vendors move at once like this, they're not following a trend. They're fighting over who gets to define it.

Context graphs vs knowledge graphs: the truth layer and the operational layer

The difference is simpler than the launch posts make it sound: a knowledge graph models what's true, and a context graph models how work used those truths. Truth layer versus operational layer. Everything else in the vendor posts is a variation on that split.

The receipts from our own lab

We didn't buy a context engine when our agent layer needed context. We wrote the rules into the vocabulary and let the graph grow from the schema we already had. Today that layer runs 46 knowledge cards, 154 nodes, and 170 edges, every one with a source and a date, and when a fact changes, every answer built on it updates. The full receipts are published, including the eleven-day disagreement that started the whole build. The governance rules are the part the new wave keeps rediscovering: an answer that cannot name its nodes is not an answer, and a role change without a source is a rumor.

Why the wave matters anyway

Steel the wave honestly and it does real work. AllegroGraph's GraphTalker packages persistence, reasoning, and audit into an agentic layer. Atlassian's context graph is about knowledge hygiene: deduplication, staleness flags, contradiction resolution before content reaches the graph. Those are operational problems every enterprise has, and packaged engines will solve them faster than internal teams can build them. Gartner projected back in 2021 that graph technologies would be used in 80 percent of data and analytics innovations by 2025. Adoption lagged. A wave of packaged context layers, priced and sold like products, is how a lagging projection becomes a default.

How to judge one

The buyer's filter has not changed since the AI SDR governance checklist, plus three questions for this new category. Where does the context live: your infrastructure, or the vendor's? Who defines the vocabulary: you, or whatever their product schema says a customer is this quarter? And what exports when you leave: your graph, your rules, and your decision lineage in a format you can reload, or a proprietary pile? A vendor who answers all three with your name on the infrastructure is selling you a context layer. A vendor who answers with their platform is selling you a subscription to your own judgment.

• If your agents disagree about facts: you need the ontology and the knowledge graph underneath, not a newer engine on top.
• If your agents disagree about judgment: exceptions, precedents, who decided and when, then the context layer earns its name.
• If your agents disagree and nobody can say which facts they used: no graph of any kind will save the deployment.

Buy for the failure you actually have. The label on the box is the wave's business, and the vocabulary underneath is yours to own.

Go deeper: Alfred case study