The human brain naturally uses experience to infer how entities, attributes, and events relate to each other, abilities that are essential for both episodic and semantic memory. Yet, the learning and inference processes that allow the brain to build relational memory, across multiple neural circuits and time, remain poorly specified. In this talk, I will discuss two sets of findings using human fMRI and temporal relation learning. In the first, I use computational theories of learning as precise but competing hypotheses for how the brain uses noisy experiential data to infer the strength of relations among visual events and to build relational memory. Results reveal a surprising computational continuity between memory formation and the mechanisms of reinforcement learning and classical conditioning. Furthermore, neural and behavioral data jointly suggest that the brain supports multiple strategies simultaneously, and raises the further question of how strategies are selected. In the second finding, I shed light on how the brain integrates diverse experiences to build knowledge of shared relational structure, in both classically episodic and semantic brain areas across two time-points. I find that entorhinal cortex, in particular, plays a key role in rapid relational integration that complements the role of semantic areas in slower integrative processes. I will raise a number of open future questions regarding the neural mechanisms of relational memory.