Adaptive processing of predictions and prediction errors in hippocampal subregions

The hippocampus is central to generating memory-based predictions and responding when reality matches or violates these predictions. Here, we examine fMRI research of hippocampal subregions in humans and identify distinct putative mechanisms underlying such predictive processing. Specifically, hippocampal regions CA3 and CA1 may both support predictive processing, but at different stages of learning. Emerging evidence supports the view that CA3 can represent memory-based predictions by pattern-completing the prediction from a cue after only minimal learning. CA1 predictive representations, however, may emerge when predictions are well learnt; CA1, thus, might serve as a 'gatekeeper' that relays established predictions to downstream regions. Through pattern separation, the dentate gyrus may support fine-grained, immediate predictions that distinguish similar events. Importantly, hippocampal responses to prediction matches versus errors depend on goal relevance. Goal-directed retrieval enhances CA1 responses to prediction matches, whereas in the absence of explicit retrieval goals, prediction errors lead to high CA1 activation through activating non-overlapping neural populations. Finally, we review evidence that prediction errors lead to both neural integration and separation, offering distinct routes for memory update. Separation versus integration might occur in parallel in different hippocampal subregions, impacted by prediction strength and neuromodulatory systems. Together, this review proposes potential mechanisms by which hippocampal subregions promote adaptive predictive processing and offers gaps and directions for future research. This article is part of the theme issue 'The role of hippocampal predictions in cognition: bridging perception and memory'.