To update or to separate: Neural signatures and consequences of latent cause inference in episodic memory

Episodic memories are leveraged to predict upcoming input. Latent Cause Theory (Gershman et al., 2017) predicts that after moderate prediction errors, existing memories will be updated by integrating new information, while after large prediction errors, a new latent cause is inferred, and the unpredicted event is encoded as a separate trace. To date, no study has directly compared memory updating with new latent cause inference in neural signatures and representational consequences. This study used naturalistic dialogues, being either coherent within the same latent cause or involving an unpredicted change of topic, causing new latent cause inference. Additionally, it was varied whether dialogues were familiar or novel, with only the former allowing the formation of high-precision predictions. While all prediction errors engaged a core network of inferior frontal and superior temporal regions, as well as Crus 2, updating existing memories with new information additionally relied on frontal midline regions (BA9 and vmPFC). Neural substrates of new latent cause inference depended on whether a precise prediction had been possible: Low-precision predictions evoked broad processing and information gathering, while high-precision predictions instead lead to a top-down adherence to the prediction. These findings suggest that updating memories with new information necessitates higher-order frontal engagement and highlight how prediction precision shapes prediction error processing. More broadly, the findings demonstrate how episodic memory representations are reorganized in response to unpredicted events, distinguishing between updating an existing trace or by forming a new, separate one.