OpenWiki and evidence graphs boost agent accuracy across insurance documents

A synthetic library of 111 insurance documents was used to test OpenWiki as a persistent memory layer for AI agents. Without the wiki, agents achieved 10% accuracy on cross-source questions; with OpenWiki, accuracy rose to 52%. Adding an evidence graph linking wiki statements to source passages further improved accuracy to 58% and doubled citation consistency. The system reduces agent workload by avoiding the need to reconstruct cross-document relationships on every question, lowering estimated costs from $0.55 to $0.19 per task.