Growing evidence suggests that the hippocampal formation functions as a relational (or episodic) memory, and on tasks of a spatial nature, it serves as a place learning and recognition system. In this paper we develop a computational characterization of the hippocampal formation pertaining to spatial learning and localization, and show that the information processing posited for this structure closely matches an engineering tool for probabilistic spatial localization of robots: the Kalman filter. Based on this parallel, we derive expressions for information integration, propagation of covariances, and near-optimal update rules for the localization performed by our computational model.