OpenWiki lifts agent accuracy 5x on connected queries

LangChain originally built OpenWiki so coding agents would not have to rediscover architecture and code behavior on every task. Each wiki statement rests on a claim tied to supporting source lines and their hash, so when those lines change, the hash changes and affected claims can be flagged without a model call. That self-correcting mechanism is what the company positioned as the persistent memory layer for agents.

The experiment used insurance because policies and endorsements frequently reference, amend, and override one another. The corpus was made intentionally messy, mixing outdated versions, contradictions, indirect reference chains, and very long documents. LangChain reports that the wiki's accuracy and cost gains on connected questions still left evidence retrieval inconsistent, with citation overlap dropping from 0.74 on simple lookups to 0.25 on connected questions. It was that gap the evidence graph was built to close.

LangChain frames the economics as a question of reuse. Simple lookups saw search alone just as accurate and faster, and ingestion plus ongoing maintenance may outweigh retrieval savings. The tradeoff shifts, in its account, when multiple agents repeatedly answer questions spanning the same disconnected sources.