Descriptions:
Zach Blumenfeld, a research engineer at Neo4j, delivers a conference talk explaining why AI agents need more than a standard knowledge base to make good decisions — and how “context graphs” address that gap. The core distinction he draws is between systems of record (facts, entities, current state) and what context graphs add: past decision traces, causal chains, precedents, and dynamic policy information that allow an agent to act with something closer to subject matter expertise.
Blumenfeld demonstrates the concept with a financial analyst agent built on Neo4j, Claude, OpenAI embeddings, and a Next.js frontend. Given a question about a loan or account, the agent doesn’t just retrieve customer data — it also pulls historical decision traces and uses a hybrid search combining semantic similarity (via vector index) and structural similarity (via graph embeddings on the Neo4j graph itself) to find precedents that match the current situation in both meaning and causal structure. The result is an agent that can recommend “reject” or “approve” and explain why, grounded in prior decisions rather than just current facts.
The talk also introduces a scaffolding CLI tool — `uvx create-context-graph` — that generates a full-stack application with a graph backend, frontend, and seeded demo data in a single command, supporting frameworks including Pydantic AI. Blumenfeld notes integrations with GitHub, Notion, Jira, and Slack for importing real organizational data, making context graphs accessible beyond synthetic demos. All code is open source and linked at the end of the presentation.
📺 Source: AI Engineer · Published May 29, 2026
🏷️ Format: Deep Dive
