Descriptions:
At the AI Engineer conference, Neo4j’s Andreas Kollegger and Zaid Zaim present ‘context graphs’ — a framework for making AI agents not just knowledgeable but genuinely decision-aware. The core argument: standard RAG pipelines give agents access to facts, but they lack the encoded policies and rules that explain *why* an agent should take a particular action. Context graphs, built on Neo4j’s graph database, address this gap by storing three layers of memory simultaneously: short-term memory (conversation state and history), long-term memory (organizations, people, and entities), and reasoning memory (policies and rules governing agent decisions).
The presenters walk through a live implementation using Neo4j’s text-to-Cypher tool, which translates natural language queries into graph traversals, and demonstrate how a decision-making workflow can be built on top of this memory structure. The framework is designed to integrate with orchestration tools like LangGraph and applies across domains — from financial services eligibility checks to medical guidance — wherever agents need to justify their actions against predefined organizational constraints.
The talk directly addresses the scaling challenge of multi-agent systems: as specialized agents proliferate and collaborate, they need a shared, auditable understanding of both the knowledge domain and the rules governing action. The presenters argue that embedding this ‘why’ layer into a graph structure — rather than relying on prompt context alone — is the key architectural move that separates brittle agents from trustworthy, production-ready ones.
📺 Source: AI Engineer · Published May 28, 2026
🏷️ Format: Deep Dive
