Turning Agent Memory Into Skills That Work — Will Lyon, Neo4j
Neo4j proposes replacing vector search agent memory with typed knowledge graphs for coherent, traceable context
“remembering is not exactly actionable knowledge”
Will Lyon of Neo4j argues that current agent memory—embedding text and retrieving similar chunks—leaves agents in an 'amnesia cycle' because raw retrieval is not actionable knowledge. He proposes 'context graphs': a connected graph of short-term, long-term, and logical memory built by extracting entities and relationships from agent messages into a shared ontology. The approach addresses a real gap in agentic systems by giving agents coherent, typed, and traceable knowledge rather than decontextualized text fragments.