Context Graphs for Explainable, Decision-Aware AI Agents — Andreas Kollegger & Zaid Zaim, Neo4j

TL;DR

• 1

  Context graphs add the missing 'why' — Zaid Zaim argues agents already have language, reasoning, and tools, but still need graphs that encode rules, policies, and decision rationale so they can justify actions instead of just retrieving knowledge.

• 2

  Memory is split into short-term, long-term, and reasoning layers — Neo4j’s model stores conversation state in short-term memory, people/org/context in long-term memory, and policy-based reasoning in a graph so future decisions can be grounded in precedent.

• 3

  Bad agent decisions show up in ordinary life first — Andreas Kollegger’s Red Bull example makes the point vividly: an agent may correctly reorder drinks but still fail if it doesn’t know rent is due, which exposes the gap between task completion and sound judgment.

• 4

  Decision-making should be an explicit workflow, not a prompt tweak — Kollegger proposes a framework that starts with local context and causality, adds global history plus hard and soft business rules, then runs risk-value analysis before any action is taken.

• 5

  Reference class validation matters when averages can kill you — In his medical example, the '99% correct' drug becomes fatal for the 1%, showing why agents must identify which population or case they’re actually dealing with instead of leaning on generic statistical defaults.

• 6

  The right output is often escalation, not action — Kollegger recommends separating proposal generation from final authority, so an agent can rank options when appropriate or hand off to a human or higher-privilege agent when certainty or authority is missing.