The benefit of being wrong: How prediction error size guides the reshaping of episodic memories

Episodic memories are not static – they shift and reshape as our surroundings evolve. One powerful mechanism for change are prediction errors, which arise when predictions about what is going to happen next do not match the actual input. This study investigated how the size of prediction errors – arising from predictions based on episodic memories – affects recognition memory and neural memory representations. In an fMRI experiment, participants listened to a series of naturalistic dialogues. In a later session, a critical segment of the dialogue was altered to introduce a mismatch and thus evoke a prediction error. Importantly, larger prediction errors were linked to increased recognition memory for the original and mismatching targets, and better source memory for the mismatching targets. Representational similarity analysis revealed that larger prediction errors were also associated with stronger reinstatement of the original version during mismatching (unpredicted) input, which promoted memory for both the old and the new version. Additionally, larger prediction errors enhanced the long-term representational stability of the original memory. We argue that these results support the idea that stronger episodic prediction errors lead to a more distinct encoding of new information, which benefits the recognition of both old and new information. This could be achieved by a pattern completion mechanism in which old information is reinstated during mismatching new input.