CauProphet preprint reports gains from agents retaining causal knowledge

Researchers report that CauProphet, an AI agent designed to retain and revise causal knowledge, improved prediction accuracy from 21.3% to 28.6% in dynamic real-world settings. The preprint also reports a 25.4% improvement over a baseline agent in mean rank correlation with next-day financial returns. Rather than retrieving only past records or summaries, the agent combines retained knowledge with current evidence to represent causal factors and possible actions. It updates that representation as observations arrive. The financial result measures correlation with returns, rather than investment profits.