The recent reawakened interest in `neural' networks begs the question of their relevance to the analysis of real nervous systems. Network models have been criticized for the lack of realism of their individual components, and because the architectures required by some neural-network algorithms do not seem to exist in real nervous systems. In three related papers published in the 1970s, David Marr proposed that the cerebellum, the neocortex and the hippocampus each acts as a memorizing device. These theories were intended to satisfy the biological constraints, but in computational terms they are underdetermined. In this paper we reassess Marr's theory of the hippocampus as a temporary memory store. We give a complete computational account of the theory and we show that Marr's computational arguments do not sufficiently constrain his choice of model. We discuss Marr's specific model of temporary memory with reference to the neurophysiology and neuroanatomy of the mammalian hippocampus. Our analysis is supported by simulation studies done on various memory models built according to the principles advocated by Marr.