While the hippocampus is known to encode physical position through place cells, substantial evidence indicates that hippocampal neurons also respond to a variety of other non-spatial sensory information. We explore the hypothesis that these representations emerge through generic computations implemented by the hippocampus that actively learn from the statistical structure of experience. Theoretical work has shown that place cell coding arises as a special case during spatial exploration in networks trained to learn sparse, compressed representations, due to the local correlations present in the sensory content of nearby locations. In computational models and experiments, we assess the impact of local correlations and environmental complexity on place fields and memory-guided behavior. We simulate how the local smoothness and compressibility of input features can systematically change the structure of the learned representations, so that the tiling properties of place fields optimally capture the geometry of the sensory manifold. In parallel, we exploit algorithms for procedural texture generation to generate parameterizable virtual environments that precisely control the complexity of visual experience in experiments. In ongoing work, we measure how the compressibility of the sensory environment controls the geometry of the learned neural representations via 2-photon imaging in behaving mice and determines the efficacy of pattern separation in human behavior. Our work refines the computational role of the hippocampus in memory formation, where adaptive statistical learning isolates the shared components of experiences for efficient storage and pattern separation.