Our environment is replete with statistical structure and similar cause-effect relationships hold across related experiences. By extracting knowledge about the relationships between items in the world, the brain can therefore predict states and reinforcements that were never directly experienced. In physical space, the hippocampal-entorhinal system organises statistical regularities between landmarks in a cognitive map, which provides a coordinate system that enables inferences about spatial relationships that were never directly experienced. In this talk, I demonstrate that a similar map-like organisation of knowledge can also be observed for discrete relationships between objects that are entirely non-spatial, suggesting that the same codes may also organise other dimensions of our experiences. When subjects need to flexibly switch between cognitive maps characterised by the same underlying structure, but a different distribution of stimuli, structural knowledge is abstracted away from sensory representations in the medial prefrontal cortex over time. Such a separation of structure from stimulus representations may facilitate the generalisation of knowledge across sensory environments and thereby accelerate learning in novel situations. Furthermore, I will talk about a recent study combining a virtual reality task with computational modeling and functional magnetic resonance imaging to understand our ability to use maps for generalisation. Together, these studies suggest potential mechanisms underlying the remarkable human ability to draw accurate inferences from little data.