Descriptions:
Stephen Chin, head of developer relations at Neo4j, delivered a conference talk at AI Engineer presenting CrabRAG — a graph-based memory architecture designed to fix the persistent memory limitations that plague modern AI agents. Using the mascot ‘Crab D’ as a running metaphor, Chin illustrates how current agent frameworks, including Claude Code, rely on flat markdown files for memory storage. This approach forces agents to load upward of 100,000 tokens of context per round, repetitively injecting everything in hopes that something relevant surfaces — a strategy that breaks down at scale.
Chin walks through the memory evolution from markdown files to vector databases to graph databases, explaining the tradeoffs at each step. Vector similarity search improves retrieval over flat files, but similarity in vector space is not the same as actual relationships. For complex multi-hop reasoning tasks, vector-only approaches hallucinate and fail — particularly in scenarios like the digital twin demo he runs live on his home lab using Claude as the agent. Graph databases like Neo4j solve this by storing entity relationships as first-class objects, enabling precise, explainable, and auditable retrieval.
The CrabRAG architecture combines both approaches: vector search identifies seed nodes to begin graph traversal, then graph search pulls nearest neighbors ranked by relational proximity. The talk also references the Hermes agent framework and an arXiv paper from Neo4j on graph-based skill selection, situating CrabRAG in the broader conversation about how agents can build durable, structured memory that persists across sessions without context-window bloat.
📺 Source: AI Engineer · Published July 22, 2026
🏷️ Format: Deep Dive
