Structured Linked Data as a Memory Layer for Agent-Orchestrated Retrieval

A team of researchers including Andrea Volpini, Elie Raad, Beatrice Gamba, and David Riccitelli has published groundbreaking research demonstrating how structured linked data can serve as a powerful memory layer for AI agents. Their paper, "Structured Linked Data as a Memory Layer for Agent-Orchestrated Retrieval," addresses a key limitation in current RAG (retrieval-augmented generation) systems that typically treat documents as flat text, ignoring valuable structured metadata and relationships that knowledge graphs provide.

Using Google's Vertex AI Vector Search 2.0 for retrieval and the Google Agent Development Kit (ADK) for agentic reasoning, the team conducted controlled experiments across four practical domains: editorial, legal, travel, and e-commerce. They tested seven different conditions comparing document representations including plain HTML, HTML with JSON-LD markup, and their enhanced agent-optimized entity page format. The enhanced format incorporates schema.org-style agent instructions, breadcrumbs, and neural search capabilities.

The results were significant: while JSON-LD markup alone provided only modest improvements, the enhanced entity page format achieved substantial gains of +29.6% accuracy improvement for standard RAG and +29.8% for the full agentic pipeline with multi-hop link traversal. The Enhanced+ variant, which added richer navigational affordances and entity interlinking, achieved the highest absolute scores with 4.85/5 accuracy and 4.55/5 completeness ratings.

Importantly, the researchers have released their complete dataset, evaluation framework, and enhanced entity page templates to support reproducibility and further development. This research provides concrete evidence that properly structured data can dramatically improve AI agent performance, moving beyond traditional flat-text approaches to leverage the full power of linked data relationships.