Researchers have proposed a new AI framework called ForeDreamer for predicting future events on the open web. The system separates memory into two types: factual memory, which stores question-specific evidence, and experiential memory, which accumulates agent experience across forecasting episodes. A main agent handles search and prediction, while a subagent processes search results to convert them into structured factual memory. ForeDreamer evolves over time by improving both forecasting decisions and factual-memory construction. Experiments on two datasets demonstrate the effectiveness of this new framework. --- Why it matters: This matters because current AI systems struggle with open-web forecasting due to noisy and incomplete evidence. ForeDreamer's ability to distill reliable signals from web data could improve prediction accuracy in critical applications such as finance, weather forecasting, or emergency response planning. Source: https://arxiv.org/abs/2608.20920