Why your agents need decision traces, not just documents — Zach Blumenfeld, Neo4j

TL;DR

• 1

  Context graphs go beyond RAG — Zach frames the shift as moving from systems that answer questions correctly with customer data, transactions, and policies to systems that can recommend “accept or reject, and why” using past decision traces and precedents.

• 2

  Graph embeddings make prior decisions searchable — Neo4j embeds not just text but connected decision-trace structures, so an agent can find similar past cases by semantic similarity and graph structure, which Zach says is hard to recover if everything stays buried in documents.

• 3

  The financial-risk demo shows the payoff — In his example, the agent retrieves Jessica’s profile, pulls prior decision traces, runs a hybrid precedent search, and returns a reject recommendation with risk factors instead of a generic summary.

• 4

  A new uvx create-context-graph tool bootstraps the whole stack — Zach shows a one-line command that scaffolds the frontend, backend, graph schema, demo data, and framework integration, with support for Pydantic AI, OpenAI, LangGraph, CrewAI, Strands, and Google ADK.

• 5

  Neo4j’s memory model is three-part: short-term, long-term, and reasoning — The underlying neo4j-agent-memory package stores conversation history, extracted entities that persist over time, and the reasoning traces that make context graphs useful for agent decisions.

• 6

  This is powerful but still early — Zach is candid that writing new decision traces, adding timestamps, and attaching quality or sentiment scores are still evolving parts of the project, even though the code and demos are already open source.