Beyond Vector Memory: Building Temporal Graph Memory for Conversational Agents

Session track: AI Engineering

Session time: 08:30 am - 09:00 am UTC November 12

Session description:

Standard agent memory relies on semantic vector stores that treat past interactions as a flat list of documents. However, vectors have no inherent concept of identity, relationships, or time. When a user says "I used to work at Apple, but last week I started at Neo4j," a vector retriever pulls both snippets as equally relevant, causing the agent to hallucinate. Standard memory fails to handle changing real-world states, leaving agents unable to track historical contexts. The presenter will show how to solve the limits of vector memory by constructing a temporal, agentic knowledge graph using Neo4j. You will explore how to represent conversational episodes as graph nodes and dynamically extract entities and relationships. The speaker will demonstrate how to design a graph schema featuring timestamped edges, allowing the agent to perform graph traversals that query "current facts" while archiving past events as "history, not noise." By attending, you will learn how to design an evolving memory graph that tracks dynamic facts over time. The speaker will share Python code and Cypher queries demonstrating how to execute hybrid retrievals (combining vector search with graph traversals) and update temporal relationships dynamically during agent execution. You will walk away with a production-ready framework to build agents that possess true, evolving long-term memory.