Road to NODES 2026 | Actionable Knowledge With Context Graphs & Agent Memory

This workshop is designed to show you why most agent memory stops at recall - you embed the transcript, retrieve the nearest chunks, and hope the model reconstructs what it already figured out last week. It usually doesn't, because the conversation is retrievable, but the knowledge was never structured in the first place. You will close that gap by working hands-on against a running Neo4j Agent Memory Service (NAMS) instance, building a memory layer that holds three distinct kinds of memory in one graph: short-term conversation history, long-term entities and relationships modeled with POLE+O, and reasoning memory that captures decision traces and tool-call provenance as first-class nodes. The workshop includes hands-on exercises centered on the two hardest problems in that stack. You will author a domain ontology and import an existing model from Arrows, Neo4j Data Importer, RDF/OWL, GraphQL, or LinkML rather than starting from a blank canvas, then dry-run it against sample text with live extraction preview and iterate until the graph you get is the graph you designed. You will work through validation modes, PII handling, allowed relationship types, the extractor's pending-types queue, the entity-resolution review queue, and versioned migrations against data that's already landed. From there, you will instrument an agent loop so every reasoning step and tool call is recorded with provenance, watch observations and reflections derive from those traces, and query the graph to distill reusable skills - answering the question that separates an agent with memory from an agent with experience: when I faced a problem shaped like this before, what actually worked? By the end of the session, you will have a running context graph, an ontology you authored and validated yourself, and an agent that learns from its own reasoning rather than re-deriving what it already knows. Bring a laptop - you'll build the whole loop, and verify every step with read-only Cypher rather than trusting an API response. Neo4j Agent Memory is a Neo4j Labs project, actively developed and community-supported; expect a few sharp edges, and we'll point them out as we hit them.