Hippocampus responds to mismatches with predictions based on episodic memories but not generalised knowledge

Prediction errors drive learning by signalling a mismatch between our expectations and reality. The hippocampus plays a key role in mismatch detection, but it is not known what information the hippocampus uses to form expectations. Here we show that the human hippocampus bases its expectations on episodic memories and not generalised schematic knowledge. Across three fMRI experiments, we demonstrate that the hippocampus is selectively engaged by mismatches with episodic memories of specific events. In contrast, mismatches with schematic knowledge activate regions of the Semantic Control and Multiple Demand Networks, as well as subcortical regions involved in prediction error signalling. Notably, mismatches with episodic memories also engage the Default Mode Network. Overall, our findings provide direct support for some models of learning via mismatch detection and refute models that propose the hippocampus plays a wider role as a more generalised mismatch detector.