Like several other groups, we have proposed that the CA3 region of the hippocampus uses the NMDA channels in recurrent collaterals to store sequences of events (places) and thereby serve as a heteroassociative network. We have specifically argued that the readout of these sequences at a rate of one memory per gamma cycle can account quantitatively for the "phase-advance" of hippocampal place cells. One problem with this proposal is that the phase-advance is also observed in dentate granule cells. We now offer a possible answer to this problem. Theoretical work has suggested that the best method for predicting sequences is to have each step-wise prediction "cleaned up" by an autoassociative network before it serves as a cue in the next step of sequence prediction. We propose that the dentate-hilar region is the autoassociative network that accomplishes this task. Two anatomical findings are relevant. First, it is known that CA3 axons have a branch that turns backward and innervates cells in the dentate/hilar region. Second, although dentate granule cells do not have recurrent axons as required in the simplest autoassociative networks, they do innervate hilar mossy cells. These in turn, make recurrent connections onto granulate cells. This disynaptic circuit could allow the dentate/hilar region to perform autoassociative function. The interaction of an autoassociative network in the dentate/hilar region and a heteroassociatve CA3 network could produce accurate sequence recall.