LangChain: graph-based agent memory lifts connected-question accuracy 27%, cuts cost 58%

JeremyCMorgan · x · 2026-11-06

LangChain tested OpenWiki, its open-source wiki compiler, as a persistent memory layer for agents:

On questions spanning multiple sources, an agent using OpenWiki as memory was 5x more accurate than one using grep tools alone—compiled memory preserves domain knowledge and relationships agents would otherwise reconstruct per question.
Turning citations into an explicit evidence graph that agents can follow doubled citation consistency, improved accuracy by 27%, and cut cost per correct answer by 58% vs. wiki alone.
The dividing line: graphs help when answers require relationships, not retrieval. On simple lookups the memory-less agent was equally accurate and faster.

Compiled memory suits repeated cross-source reasoning (e.g., banking/insurance policies that amend each other); skip it for simple retrieval.

More from coding & agent