Descriptions:
Stephen Chin, head of developer relations at Neo4j, presents a technical case for knowledge graphs as the context layer that standard vector RAG architectures are missing. The talk, delivered at AI Engineer Singapore, opens with a framing around enterprise AI’s core problem: agents making decisions from siloed, disconnected data sources that lack relationship context.
Chin walks through a concrete healthcare example — retrieving a care plan for a named patient — comparing outputs from a baseline LLM, a vector RAG system, and a graph-grounded system. The graph-grounded response is notably more specific and clinically accurate, surfacing patient history, prior diagnoses, and personalized treatment details the other approaches miss. He then introduces the broader concept of a “context graph,” now recognized by Gartner on the AI Hype Cycle, which unifies short-term memory, long-term memory, and agent reasoning traces into a single queryable Neo4j graph structure.
A practical highlight is the Neo4j Agent Memory open-source package, which implements all three memory types in graph form. Chin explains how LLMs can generate Cypher (Neo4j’s query language) natively, how graph algorithms like Louvain community detection help navigate complex knowledge structures performantly, and how stored reasoning traces enable compliance auditing and improved future decisions. The talk is aimed at engineers building production RAG pipelines or multi-step agent systems who need explainability and richer retrieval than flat vector search can provide.
📺 Source: AI Engineer · Published May 16, 2026
🏷️ Format: Deep Dive
