Why your agents need decision traces, not just documents — Zach Blumenfeld, Neo4j
Summary
Zach Blumenfeld, a research engineer at Neo4j (a graph intelligence company), presents the concept of context graphs and explains why AI agents need decision traces in addition to conventional knowledge bases. The talk was inspired by Foundation Capital's work in December 2025 around decision traces and reasoning to help agents make better decisions.
Blumenfeld draws a clear distinction between a knowledge base and a context graph. A knowledge base powers standard RAG by helping an agent answer questions correctly — it supplies facts, entities, and current state. A context graph extends this with the information required to make better decisions: past decision traces, precedents, causal chains, and dynamic context about why decisions were made. The canonical example is a financial analyst agent evaluating a loan or transaction request. With only a knowledge base, the agent might return a risk score and suggest a review. With a context graph, it can give a definitive accept/reject recommendation grounded in structurally similar past decisions and the reasoning behind them.
The context graph data model has three layers: short-term memory (conversation history, session context), long-term memory (extracted entities that persist and are resolved over time), and reasoning traces (the decision traces themselves). A key technical innovation is graph embeddings: just as text is embedded into vectors for semantic search, decision trace subgraphs are embedded into vectors, enabling lookup of structurally similar past decisions alongside semantically similar ones. This hybrid search — combining vector similarity on text with structural similarity on graph topology — is what surfaces the most relevant precedents for a new decision.
Entity extraction is handled by the neo4j-agent-memory package, which implements a staged NLP pipeline: spaCy for initial extraction, GLiNER for more advanced named entity recognition, and an LLM as a fallback. A separate merging, deduplication, and enrichment stage resolves entities across sessions. The package also manages the transition from short-term conversation data into long-term persistent entities.
The fastest path to building with context graphs is the uvx create-context-graph CLI command. A single command — specifying app name, domain, agent framework, and whether to use demo data — scaffolds a full-stack application with a Next.js frontend, a backend, demo fixture data, an MCP server, and multi-turn conversation support. The tool ships with 22 built-in domains (including healthcare and financial services) and can also generate a custom ontology from a user-described domain. Data connectors are available for GitHub, Notion, Jira, and Slack. Supported agent frameworks include Pydantic AI, LangGraph, Crew, Strands, Google ADK, and OpenAI. The demo shown used Claude as the agent runtime, OpenAI embeddings, and Neo4j with GDS (Graph Data Science) for graph embeddings.
Graph Data Science (GDS) is the Neo4j component used to compute graph embeddings. Once decision traces are embedded, a vector index enables similarity search: querying for fraud rejection cases, for example, retrieves structurally similar prior decisions just as vector search retrieves semantically similar documents. This capability is described as particularly hard to replicate with a flat document store.
Blumenfeld acknowledges the project is early-stage and open source, inviting community contributions. Open questions discussed in the Q&A include automatic decision trace capture (currently requiring explicit user prompting), quality scoring of decisions, and handling structured vs. unstructured input data when generating ontologies.
Key Takeaways
- •A knowledge base tells agents what is true; a context graph tells agents what to decide and why, by supplying past decision traces and precedents alongside factual data.
- •Graph embeddings extend vector search to structural similarity: embed decision trace subgraphs into vectors so that structurally similar past decisions can be retrieved alongside semantically similar ones.
- •Run uvx create-context-graph to scaffold a complete full-stack context graph application — backend, Next.js frontend, demo data, and MCP server — in a single terminal command.
- •The neo4j-agent-memory package handles entity extraction via a three-stage pipeline (spaCy → GLiNER → LLM fallback) with built-in merging and deduplication, so short-term conversation data transitions cleanly into persistent long-term entities.
- •Context graphs require three memory layers: short-term (conversation/session), long-term (resolved entities), and reasoning (decision traces); all three are necessary for agents to act with subject-matter expertise.
- •22 built-in domains (healthcare, financial services, etc.) are available out of the box, and the tool can generate a custom graph ontology from a plain-language domain description.
- •Real-world data can be imported via connectors for GitHub, Slack, Notion, and Jira — no need to rely on synthetic demo data.
- •The create-context-graph scaffold supports Pydantic AI, LangGraph, Crew, Strands, Google ADK, and OpenAI frameworks, so teams can stay within their existing agent stack.
- •Hybrid search — vector similarity on text combined with structural similarity on graph topology — surfaces the most relevant precedents and is difficult to replicate with flat document stores.
- •The project is open source and early-stage; automatic decision trace capture and quality scoring of decisions are known gaps being actively worked on.
Transcript
So, my name is Zach. Uh I work for Neo4j. We're a graph intelligence company. You can think of us like a knowledge layer graph database at the core. We help connect and resolve information so AI systems can be a little bit more accurate and explainable. Um I used to work in technical marketing. I very recently I'm going back to my AI and machine learning roots. Uh I'm transferred over to a to a research engineering role. So, I'm going to talk to you today about context graphs. How many people in...
Related Sessions
