William Lyon is a senior product manager at Neo4j developing graph-based infrastructure that helps AI agents retain knowledge, explain decisions, and transfer experience between systems. The author of Full Stack GraphQL Applications, he has shaped developer tooling spanning graph databases, GraphQL, geospatial analytics, GraphRAG, and agent memory.
Lyon earned a master’s degree in computer science from the University of Montana and worked as a startup software engineer before joining Neo4j. An early Google Summer of Code project added spatial capabilities to Cypher, establishing his interest in graph-backed geographic data.
At Neo4j, he worked in developer relations and advocacy. In 2017, he and Michael Hunger developed an early Neo4j–GraphQL integration that generated graph-backed APIs from GraphQL schemas. He subsequently championed GRANDstack, bringing together GraphQL, React, Apollo, and Neo4j, and published his practical guide to full-stack GraphQL applications in 2022.
In 2023, he joined Wherobots to lead developer relations, working on geospatial analytics and Apache Sedona. He then joined Hypermode, where he worked with Dgraph, knowledge graphs, and AI agents. At AI Engineer World’s Fair 2025, his knowledge-graph mullet described combining intuitive property-graph interfaces with RDF-backed storage. Later that year, he returned to Neo4j’s product team to focus on graph-based agent memory.
- GraphRAG beyond vector search: Similarity search locates an initial passage; graph traversal connects it to related people, organizations, places, topics, and documents. Geospatial indexes can provide additional entry points into the same knowledge graph.
- Property graphs over RDF triple storage: Applications expose familiar nodes, relationships, and traversals while representing data underneath as subject-predicate-object triples. Predicate-oriented posting lists accelerate traversal, and Dgraph’s GraphQL-inspired DQL makes querying approachable.
- Graph-native agent memory: Lyon separates conversational state, long-term knowledge about entities and relationships, and reasoning traces recording decisions, tool calls, outcomes, and provenance. The Neo4j Labs agent-memory project brings those layers together so agents can preserve context and make their behavior inspectable.
- Portable agent skills: His work on skill distillation converts relevant domain knowledge, recorded actions, and outcomes into reusable procedural instructions, allowing another agent to inherit a process without reconstructing it from conversation history.
