Context Graph Beats Vector RAG for Multi-Agent Memory

A structured context graph achieved 88.9% accuracy at 26.9 tokens per query on multi-agent conversations, outperforming a raw history dump (61.1% at 490.9 tokens) and a vector-only RAG pipeline (50.0% at 75.9 tokens). The graph stores facts as entities and relationships instead of text chunks.