Aktiver — semantic AI platform Deterministic · traceable · behind your firewall
Insights / Knowledge graphs

Neo4j launches Context Engine, betting that agents need an ontology

On 20 October 2026 Neo4j announced Context Engine, a set of tools for building a shared knowledge layer that AI agents can reason over. At its centre is a business ontology: a predefined six-layer graph model for mapping an organisation’s structure, processes, data and AI assets. The product is in private preview, with Neo4j inviting companies to become design partners, according to the company’s announcement.

The launch is one more sign of a shift that has been building all year: vendors that once sold graph databases, data platforms or search are now selling “context” for agents, and the word ontology has moved from academic papers into product launches.

/ 01What Neo4j announced

Neo4j describes four parts:

  • Business Ontology, the predefined six-layer model for mapping the organisation into the graph.
  • Context Build and Runtime, which builds the ontology from existing sources such as Databricks and Snowflake and provides GraphRAG retrievers, command-line tools and Python packages.
  • Context Studio, a control plane with visual ontology and workflow builders, and an Agent Setup feature for configuring agents, Model Context Protocol servers and developer tools.
  • Memory and Learning, including decision tracking, which logs agent reasoning and operational paths as a lasting record, and knowledge distillation, which uses graph algorithms to condense repeated events into reusable knowledge.

Sudhir Hasbe, Neo4j’s president of technology and chief product officer, told TechTarget the product grew from watching customers struggle with terms like “account” and “supplier” that meant different things in different systems. He described the goal as “a shared knowledge layer that every agent can draw on more reliably.”

/ 02What the analysts said

William McKnight of McKnight Consulting called the release “an essential evolution for Neo4j and its customers,” and singled out graph traversal as the advantage: “targeted, deterministic traversals that provide precise, high-signal facts directly to the model.” He also named its main limitation. As built, Context Engine is read-only, without native change-data-capture connectors or write-back, and he said it must “move beyond read-only retrieval” to become an operational layer. He called for independent benchmarks of throughput, latency and cost.

/ 03Why it matters

Hasbe’s “account” and “supplier” example is the whole problem in two words. Agents that read raw tables and documents inherit every inconsistency in them. A sales system’s account is not a billing system’s account, and a language model has no reliable way to know that unless someone tells it. Retrieval over text finds passages that mention the word; it does not resolve what the word means in each place.

An ontology resolves it once. It defines the concept, records how each source system’s term maps to it, and lets agents query the resolved model instead of guessing. Neo4j’s decision tracking adds the other half: a record of what the agent did and why, which is what auditors and operators need when something goes wrong.

/ 04Predefined versus principled ontologies

There is a design choice embedded in “predefined six-layer model”. A vendor-defined structure is quick to start with and consistent across customers. The open question is how well it extends into specialised domains, clinical, intelligence, engineering, where the concepts are many and the distinctions subtle, and how easily the result can be shared with partners who use other tools.

Upper ontologies such as BFO take a different route. They define a small number of very general categories, such as objects, processes, qualities and roles, and domain ontologies are built beneath them. The benefit is interoperability: two organisations that both align to BFO can exchange data with a shared understanding of what kind of thing each entity is, and the result can be exported in open standards such as OWL and RDF. The cost is that someone has to do the alignment, which is where automation helps.

/ 05What to watch

  • General availability and what changes between preview and release.
  • Write-back. Whether the knowledge layer becomes operational, as McKnight urged, or stays a retrieval layer.
  • Benchmarks. Independent measures of accuracy and cost against vector-only retrieval.
  • Portability. Whether the business ontology can be exported in open formats and aligned to standard upper ontologies.

The broader point stands whichever tool is used: agents are only as reliable as the meaning they are given.

Aktiver’s take — Enterprise data

Neo4j is right that agents need an ontology, and decision tracking is the right companion to it. The open question is whose ontology, and whether it can leave the platform it was built in.

A predefined business model gets one team to a first graph quickly; a BFO-aligned one lets two organisations share a graph. When “supplier” in procurement, “vendor” in accounts payable and “manufacturer” in quality are mapped as roles borne by the same organisation entity, an agent can answer which suppliers with open quality findings were paid this quarter without guessing that three tables mean the same thing. Each mapping cites the system and field it came from, so a wrong join can be traced and corrected by the subject-matter expert who owns that data. Our Ingestion Hub builds that kind of model from PDFs, SQL, CSV and JSON, lets SMEs refine it, and exports OWL and RDF, so the meaning stays portable across graph stores.

How we’d model it

SupplierOrganizationProcurement recordAgent decisionOntology pathBusiness ontologyBFOisRoleOfmapsTotraversesalignedToSupplierOrganizationProcurementrecordAgentdecisionOntology pathBusinessontologyBFOisRoleOfmapsTotraversesalignedTo
  • Supplier —isRoleOf→ Organization
  • Procurement record —mapsTo→ Supplier
  • Agent decision —traverses→ Ontology path
  • Business ontology —alignedTo→ BFO

See where this applies →

Go deeper

See how Aktiver turns raw data into knowledge graphs your AI can reason over — deterministic, traceable, behind your firewall.

← All insights Subscribe via RSS → Book a demo →
Keep reading