A growing body of research suggests that episodic memories are not only used to remember the past, but also to make predictions about the future. When a prediction is incorrect, a prediction error (PE) arises, which can promote encoding. Previous studies have usually established artificial episodic memories and then evoked PEs by introducing unexpected changes to the stimuli. However, this does not sufficiently take into account that individuals differ in prior knowledge, previous experiences and beliefs, which should inform the initial prediction model employed when facing a novel stimulus. We therefore tested whether initial PEs at first encoding of new material can predict memory stability even in the face of future modifications.

To this end, participants encoded naturalistic dialogues while undergoing fMRI scanning and then rated the dialogues regarding five parameters thought to influence the initial PE: autobiographical association, social consistency, everyday typicality, arousal and valence. Later, some of the dialogues were modified to varying degrees. After that, the original dialogues were played again to establish their stability after having been modified. Consistent with previous research, a post-fMRI recognition test revealed that episodes that had higher arousal during encoding were remembered better, regardless of the strength of the modification. Representational similarity analysis in the bilateral hippocampus showed that representations of episodes changed more after modification when participants had a vivid autobiographical association and especially if the episode was negative. These results show that individual, initial PEs influence memory representations and outcomes after memory modification, even when controlling for PE strength.