Predictions Shape Event-Based Episodic Memory Organization

People understand continuous experiences by segmenting them into distinct and meaningful events. Segmentation occurs at critical moments called event boundaries. Predictability is a cue for segmentation, yet it is not well-understood how different forms of predictability might affect event-based memory organization. For example, people could segment events at moments of prediction error or alternatively at moments of prediction uncertainty. However it is not known whether event boundaries that arise from such distinct prediction states could have distinct effects on memory. We investigated these ideas using an auditory narrative task in which participants segmented narratives into events. We applied a large language model to the narratives to estimate the probability distribution of the next word at each point in the story. We used these probability distributions to compute dynamic measures of prediction states from word to word. This allowed us to compare human segmentation behavior to model-derived estimates of prediction states. In both online and in-person experiments we found that event boundaries were associated with both prediction error and prediction uncertainty. In a third experiment, we found that prediction states at event boundaries shaped later memory for the narratives. Memory for events was higher and more integrated with subsequent information when event boundaries occurred at states of high prediction error. In contrast, memory for events was less integrated when event boundaries occurred at states of high prediction uncertainty. The findings suggest that people perceive event boundaries based on different prediction signals, which have distinct effects on the organization of memory.