Abstract
The hippocampus supports episodic memory by binding spatial and semantic information, yet how this information is simultaneously organized along its long axis remains debated. Gradient accounts propose a continuous shift in representational scale, from coarse coding in anterior to fine coding in posterior regions, whereas modular accounts posit discrete subregions specialized for distinct functions. Using high-resolution fMRI together with eye tracking as a readout of spatial and semantic memory during sequence learning, we directly tested these competing models. During predictable sequences, hippocampal activity continuously varied along the long axis. In contrast, modular organization emerged when sequences mismatched memory. Subregions in the anterior and posterior hippocampus were sensitive to semantic and spatial mismatches, respectively. Notably, the intermediate hippocampus was specifically sensitive to concurrent mismatches in both dimensions, but not to mismatches in either dimension alone. These content-sensitive subregions were embedded within distinct cortical networks that reorganized according to memory demands. Together, our findings reveal a dynamic hippocampal architecture that flexibly combines gradient and modular principles to simultaneously represent the spatial and semantic content that defines episodic memory.