Replay builds an efficient cognitive map offline to avoid computation online

Experiences are often fragmented. Reorganizing them into a coherent structure, or cognitive map, allows for inferences about unobserved relationships. A prominent theory suggests that offline replay is important for this reorganization of knowledge. However, testing this idea is hard because it requires tracking neural activity throughout the entire course of learning and reorganization. Here, we measured replay in humans throughout this process using Magnetoencephalography (MEG). This takes about 8 hours. We show that replay learns the cognitive map offline during rest, reducing the need for online computation during inferences. Subjects first experienced fragmented information along one-dimensional slices of a two-dimensional (2D) map. During rest, fast offline replay played trajectories in the 2D space - forming links that had never been experienced. This offline replay predicted the later emergence of a compact grid-cell-like code online. It is a generalizable schema representation across maps. During inference, we observed two different speeds of replay. Slow replay preferentially played out the immediate inference and was localized to the prefrontal cortex. Fast replay, developing later, played the background map and was associated with the medial temporal lobe. Fast replay was positively correlated with grid-cell-like representations, while slow replay was negatively correlated and linked to poorer inference performance. This indicates that the learned cognitive map reduces the need for slow, trial-specific computation online. Overall, these findings suggest that offline replay reorganizes sporadic experiences into structured knowledge, enabling efficient behavior.