Neo4j's Zach Blumenfeld on AI Agents and Decision Traces

In a recent presentation, Zach Blumenfeld from Neo4j highlighted the critical need for AI agents to possess "decision traces" rather than relying solely on documents. Blumenfeld, a research engineer at Neo4j, explained that for AI agents to be truly accurate and accountable, they need a mechanism to record and recall their decision-making processes. This goes beyond simply storing information; it involves understanding the "why" behind an agent's actions.

The Power of context graphs for AI agents

Blumenfeld emphasized that context graphs, powered by Neo4j's graph database technology, are essential for providing this crucial layer of information. These graphs help AI agents connect and resolve information, enabling them to be more accurate and make better decisions. He contrasted this with a simple "knowledge base," which primarily provides information to answer questions, whereas a context graph provides the information needed to make decisions.

Decision Traces vs. Documents

To illustrate his point, Blumenfeld presented a scenario involving a financial analyst agent. When asked to approve a credit limit increase, a traditional agent might only retrieve customer information and transaction data. However, an agent equipped with decision traces and context graphs could analyze past decisions, identify relevant precedents, and consider key risk factors like fraud flags or compliance issues. This allows the agent to not only make a decision (approve or reject) but also to provide a clear rationale for that decision, including the "why." This is vital for auditability and building trust in AI systems.

Neo4j's Agent Memory Solution

Neo4j offers a comprehensive API for AI agent memory, encompassing short-term memory (conversation history and session context), long-term memory (knowledge graphs of entities and relationships), and reasoning memory (context graphs for explainability and learning). The core of this system is the Neo4j Context Graph, which combines vector and graph traversal capabilities. This allows agents to search not only for semantic similarity in text but also for structural similarity within the graph data.

Introducing 'Create-Context-Graph'

Blumenfeld also introduced a new open-source project called "Create-Context-Graph." This interactive CLI scaffolding tool is designed to generate complete, domain-specific context graph applications. Similar to tools like `create-next-app`, it aims to simplify the process of building AI agents backed by context graph memory. Users can specify the application name, domain (with pre-built options like healthcare, finance, and real estate, and the ability to create custom domains), and the framework they wish to use, such as PydanticAI, Claude Agent SDK, or OpenAI Agents.

The tool also supports numerous SaaS data connectors, including GitHub, Slack, Jira, and Salesforce, enabling agents to tap into a wide range of data sources. The Neo4j agent memory package itself provides graph-native AI agents with Cypher-powered tools for querying entities, relationships, and decision traces, allowing call streams to appear in real-time with live progress indicators. For those using Claude, an MCP Server for Claude Desktop can optionally generate an MCP server config, allowing Claude Desktop to query the same knowledge graph as the web application.

Blumenfeld concluded by showcasing a demo of the financial analyst agent, illustrating how it uses context graphs and decision traces to arrive at a more informed recommendation, such as rejecting a credit limit increase due to identified risk factors. He also pointed to resources for those interested in exploring the technology further, including links to demos, the "create-context-graph" tool, and the Neo4j agent memory repository.