The hippocampus has been extensively studied over six decades, and in that time many good computational models have been developed to explain the role of the hippocampus in episodic memory. However, there are two crucial elements that have been largely neglected in these classic models. The first is the nature of the semantic input to the hippocampus. In many models, input information is treated either as unstructured, or with extremely simple structure (e.g. degrees of similarity within an environment or task). This fails to account for the rich semantic structure we extract from the world, and its crucial impact on memory processes. I will present evidence that episodic memory, and particularly false memory, depends crucially on the nature of the abstract semantic representations within the anterior temporal lobe. I will further show that these semantic representations are well predicted by a computational model that learns from the statistical structure of natural language. Secondly, there has been increasing evidence in recent years that the hippocampus plays a critical role in reasoning and inference, as well as memory. I will present evidence that such results can be explained by the recurrent connectivity between the hippocampus and entorhinal cortex, as predicted by a recent model of episodic inference.