A neural network model of sequential memory retrieval during free recall

Humans exhibit highly structured patterns of memory recall. For example, when recalling a previously studied sequence of items, two patterns are often observed. First, after recalling any given item, the next recall tends to be an item studied in temporal proximity to the just-recalled item. Second, items tend to be recalled in the same order they were studied. To explain these findings, a prominent computational model called Temporal Context Model (TCM) posits the existence of a slowly-drifting temporal context that mediates encoding and retrieval of item sequences. While this model uses a handcrafted linear rule for updating the temporal context, here we study if and how a dynamic neural system might learn a context-like representation to support sequential retrieval. We developed a memory model with recurrent dynamics and trained it on a free-recall task to investigate what patterns of recall emerged after learning. We found that the model had a propensity to recall items in forward order, even though our training process had no built-in bias for forward recall. Our model learns a joint representation of the temporal context and item-related information, allowing the hidden state of the network to evolve in the same way during recall as it does during encoding. Our findings provide a possible answer to how the neural system can encode temporal context information and perform sequential memory retrieval. Importantly, the enhanced flexibility of our model over TCM allows it to be employed in tasks other than free-recall such as decision making.