Enhanced memory for salient events can be driven by affective responses and by surprising outcomes. Recent work has suggested that deviations in expected feelings can influence decision making alongside a traditional learning signal, the reward prediction error. Whether a dynamic measure of “affective surprise” may serve as a learning signal that modulates the formation of lasting memories remains unclear. This study investigated how affective surprise elicited by dynamic changes in affective states influences long-term episodic memory. We introduce a novel computation for affective surprise informed by literature on prediction error-driven learning and derived from individual participants' continuous affect ratings of valence and arousal. We first reanalyzed a published dataset and then conducted an independent replication study in which participants encoded item sequences while listening to emotional music. Participants then re-listened to the music while providing continuous ratings of their felt affect. We assessed how affective surprise influenced different aspects of episodic memory after 24 h. We found that greater affective surprise at a given moment, or larger deviations from the recent history of ratings, enhanced memory for when an item occurred in a sequence. In contrast, we found inconsistent effects of affective surprise on item recognition memory across the two studies. Together, these findings suggest that affective surprise, regardless of direction, enhances the binding of items to their temporal contexts in memory. This work demonstrates that affective surprise, in particular valence-related surprise, might act as a learning signal with consequences for episodic memory. • Within-person changes in affective state may act as a learning signal and influence memory. • Novel metric for “affective surprise” captures discrepancy between previous and current affect. • Greater valence-related surprise enhances memory for items in temporal context. • Influence of affective surprise on memory replicated across two samples.