Animals can learn to repeat actions that lead to rewards (model-free learning). However, animals can also learn about environmental structure without receiving immediate and explicit rewards (latent learning), and then use this knowledge to guide future reward-seeking behavior efficiently. How the brain accomplishes this is not clear, in part due to limitations in existing behavioral tasks to study latent learning in rodents. In particular, traditional latent learning tasks suffer from their “one-shot” characteristics, such that learning is typically performed and tested in one environment, preventing repeatable neural experiments within subjects. Here, we developed a repeatable latent learning task by designing a physical maze with fixed reward locations (goals) and an easily reconfigurable structure. At the beginning of each session, mice explored the maze without receiving rewards. Once rewards became available (each indicated by a visual cue near a goal), mice instantly made optimal navigation decisions (a sequence of 5-7 left or right turns) to the cued goal, revealing latent learning without immediate reward. Given the absence of rewards and cues during exploration, a standard model-free reinforcement learning algorithm (Q-learning) could not learn this cue-guided task through latent learning. Temporary and reversible inhibition of the hippocampus reduced exploration of goal locations and caused errors in navigation decisions. Therefore, we report the development of a repeatable hippocampus-dependent spatial latent learning task in which mice learned without immediate reward to perform cue-guided navigation to goals.