Context Graphs for Explainable, Decision-Aware AI Agents — Andreas Kollegger & Zaid Zaim, Neo4j
Summary
Andreas Kollegger and Abhik (Zaid Zaim) from Neo4j present a framework for building decision-aware AI agents using context graphs — a structured approach that moves beyond knowledge retrieval to encode the 'why' behind agent actions, incorporating policies, rules, and historical precedent directly into the graph data model.
The talk opens by identifying the core gap in current agentic systems: agents are strong at language, reasoning, and creativity, but lack the contextual decision-making grounding that humans bring implicitly. The speakers distinguish between three types of agent memory — short-term (conversation and state history), long-term (organizations, people, entities), and reasoning memory (policies, rules, and prior decisions) — and argue that context graphs are the mechanism for operationalizing reasoning memory at scale.
Abhik introduces a five-stage decision-making framework designed to be implemented in tools like LangGraph or Google ADK. The stages are: (1) Frame the problem by capturing the objective, the causality that led to the current decision point, and the operational environment (e.g., purchasing vs. medical). (2) Pull in global context: prior decisions made in similar situations, and both hard rules (formally encoded business logic) and soft rules (informal guidance from Slack, Google Docs, etc.). (3) Conduct a risk-value analysis that includes reference class validation — determining which population subgroup the current case belongs to before applying statistical guidance. The canonical example given is medical prescribing: drug X is correct 99% of the time for symptom Y, but fatal for 1% of patients. Statistical reasoning alone cannot save the 1%; you must know which class the patient belongs to before acting. Additional risk factors include reversibility of the decision and the cost of being wrong. On the value side, the agent must have an explicit optimization target — minimizing cost versus maximizing a specific outcome. (4) Either act or escalate. This stage separates the analysis agent (which produces a ranked list of alternatives with pros and cons) from the action agent (which has authority to act or must escalate to a higher-authority agent or human in the loop). The talk stresses that a compartmentalized multi-agent design is preferable to monolithic agents that both analyze and act. (5) Write the full reasoning chain back into the graph. Every decision — including what was considered, what was rejected, and the final action taken — becomes a persisted node in the graph. Future agents inherit this as precedent, enabling the system to self-improve over time without retraining.
The framework is grounded in Neo4j's graph data model, where nodes represent entities (organizations, people, rules, decisions) and relationships encode the deep connections between them. The speakers demonstrate how a memory graph looks in practice: green nodes for short-term conversational state, extended nodes for long-term organizational knowledge, and reasoning nodes for policies and decision records. Cypher queries (via a text-to-Cypher tool) allow agents to traverse the graph for relevant knowledge before responding.
A recurring theme is making implicit human knowledge explicit for agents. The speakers note that much of current AI engineering practice is precisely this translation work — capturing the unstated assumptions and domain-specific heuristics that humans apply automatically. The framework is intentionally general but acknowledged to require significant domain-specific customization at each stage. The speakers invite attendees to discuss domain-specific implementations and direct them to Neo4j's Graph Academy for free online courses on context graphs.
Key Takeaways
- •Context graphs extend knowledge graphs by adding policies, rules, and prior decisions — giving agents the 'why', not just the 'what'.
- •Implement reference class validation before acting: determine which subgroup the current case belongs to rather than relying on population-level statistics (e.g., the 1% fatal drug case).
- •Always assess decision reversibility and cost of being wrong as explicit inputs to the risk analysis stage — these change the acceptable action threshold.
- •Separate the analysis agent (produces ranked alternatives with pros/cons) from the action agent (has authority to act or escalate); do not build monolithic agents that do both.
- •Define an explicit optimization target for each agent context — maximizing value and minimizing cost are not the same objective and must be stated explicitly.
- •Persist the full reasoning chain back into the graph after every decision; each resolved decision becomes precedent that future agents can query and inherit.
- •Encode both hard (formal business logic) and soft (informal Slack/Docs guidance) rules in the global context layer of the decision framework.
- •Frame the decision environment explicitly for agents — medical context and purchasing context require very different risk tolerances, and agents cannot infer this without being told.
- •Use a human-in-the-loop or higher-authority agent escalation path whenever the acting agent lacks certainty or authority — deferral is a valid and important system state.
- •The five-stage framework (frame → global context → risk-value analysis → act or escalate → record reasoning) can be implemented in LangGraph, Google ADK, or custom skill-based architectures.
Transcript
[music] >> Hello, good evening developers. How is everyone doing? Welcome to AI engineer and welcome to the Neo4j context graph session number two. Excited to be here together with my peer Abhik to to share with you more about context graphs, how they can help AI agents to become more decision aware. So, one of the big challenges that we are trying to solve these days in the AI era is basically um using knowledge graphs to unlock AI where we want to use graphs to fill the gap of knowledge, give ...
Related Sessions
