Connecting the Dots with Context Graphs — Stephen Chin, Neo4j

Stephen Chin, head of developer relations at Neo4j, presents the case for context graphs as a foundational architecture for enterprise AI agents at AI Engineer Europe 2026. The talk runs approximately 17 minutes and is structured around a core diagnostic: vector RAG alone fails because it loses the relational and causal context that makes retrieval meaningful.

Chin opens with a framing borrowed from The Matrix — engineers today are trapped between powerful AI tools they do not fully control and siloed enterprise knowledge systems that prevent those tools from making well-reasoned decisions. He notes that Gartner has added context graphs to the AI Hype Cycle and that Foundation Capital cited a $3 trillion startup opportunity in the space, positioning graph-based context as the next architectural frontier.

The healthcare demo grounds the argument concretely. When asked about Andre Jenkins' emphysema care plan, a baseline LLM returns generic guidance (preventing lung damage, standard protocols). A vector RAG system adds marginally more context but still produces generic advice — respiratory therapy, deep breathing, coughing exercises. A graph-grounded system, by contrast, retrieves the patient's smoking history, prior surgical operations, and connected diagnosis records, and produces specific recommendations: smoking cessation counseling, medication management, and targeted pulmonary rehabilitation. The difference is not in the data — all three systems had access to the same underlying records — but in whether traversable relationships existed to surface relevant context.

Chin then introduces a three-layer memory model for agentic systems built on knowledge graphs: short-term memory (current pipeline state, active conversation), long-term memory (cross-session history, domain entity models, business process records), and reasoning traces (the decision provenance layer that captures which tools were called, which policies applied, and why a given recommendation was made). He argues that LLMs typically return conclusions without exposing the intermediate reasoning, and that storing reasoning traces in a graph makes past decisions queryable, repeatable, and auditable — critical for compliance use cases.

The technical architecture described uses Neo4j as the graph backend with Cypher as the query language. The system combines graph traversal with vector embeddings (using OpenAI embeddings loaded into Neo4j), and applies algorithms such as fastRP for graph embeddings and the Louvain algorithm for community detection. Chin highlights that multi-hop traversal in knowledge graphs is natively performant compared to join-heavy relational approaches — a point he says is supported across multiple GraphRAG research papers.

The open-source Neo4j Agent Memory package is introduced as a reference implementation that unifies the three memory tiers into a deployable context graph structure. Two demo applications are showcased: a Lenny's Podcast knowledge graph that extracts topics, entities, and locations across episodes into a queryable graph with a map visualization; and a financial services loan decisioning application connecting a CRM, support ticket system, and internal business data via 10 MCP tools. In the financial demo, a query about loan applicant Jessica Norris surfaces a prior loan rejection, related margin trades, and fraud detection patterns — all rendered as a visible graph traversal in the Next.js UI — and the model recommends denial with explicit risk factor justification. Chin emphasizes that this explainability is what allows a human decision-maker to stand behind the recommendation.

Chin closes by pointing to Neo4j's free GraphAcademy context graph course, which spins up a free Aura cloud instance for hands-on practice, and previews deeper agentic use case sessions from colleagues Zaid and ABK following his talk.