Vercel SDK Integration with Neo4j

Session track: Modern Applications

Session time: 03:30 am - 04:00 am UTC November 12

Session description:

LLM-powered chat applications are traditionally stateless, often suffering from "amnesia" that forces them to re-derive context, leading to expensive and inefficient prompt-stuffing. This presentation explores a production-ready solution: integrating the Neo4j Agent Memory System (NAMS) with the Vercel AI SDK. We will walk through the implementation of a persistent memory layer that categorizes knowledge into short-term, long-term, and reasoning graph structures. I will demonstrate two distinct integration patterns: Provider Mode: A transparent middleware approach for seamless, "invisible" memory integration. Tools Mode: A model-driven approach that grants the agent agency over memory, providing auditability and debuggability. We will also cover best practices for using the Model Context Protocol (MCP) to bridge conversational memory with live graph data and discuss strategies for optimising token usage and memory retrieval latency.