Graph memory for LLMs is the missing layer in the enterprise AI stack. Stephen Chin, VP of Developer Relations at Neo4j, joins Alan Shimel on Techstrong TV. Furthermore, they trade notes on a year of AI conferences, the state of open weight models and why graph memory for LLMs now defines real enterprise value.

About Stephen Chin

Stephen leads developer relations at Neo4j. In addition, he has spent years as the Java man and later at JFrog on DevSecOps. Consequently, he brings a working developer view to every talk on AI, graphs and agent workflows.

Inside graph memory for LLMs

Stephen argues that frontier models look great on small data. As a result, benchmarks make them shine when a task fits inside a 200K context window. Meanwhile, real enterprise data sets crush that envelope. Therefore, LLMs need a structured knowledge layer to reason at scale, and that is where graph memory comes in.

He walks through a home lab digital twin with pgvector and GraphRAG. In addition, algorithms like k nearest neighbor and Louvain group related nodes into communities. Consequently, an agent can fetch a tight, relevant slice of the graph and hand a clean context to the LLM.

Why this matters now

Meanwhile, open weight models like Kimi K3 keep closing the gap with the top labs. Furthermore, Stephen argues that graph memory for LLMs is what separates a slick demo from a reliable enterprise system. In short, without a semantic layer, agents guess. With it, they answer.

Explore more artificial intelligence coverage and the latest Techstrong TV interviews. In addition, Stephen points newcomers to Graph Academy, Codex and Claude Code as easy entry points. He also traces graphs from Google PageRank to Panama Papers investigations across the world, and previews the next round of graph plus agent tooling coming soon.

For more information please visit neo4j.com