The Potts associative network, a neurally-informed generalization of the Hopfield model, can serve as a simplified representation of global cortical dynamics. Considering memories to be sparsely distributed patterns of local cortical activity, we have studied latching dynamics – the largely random hopping through global attractor states driven by adaptation – as a regime potentially underlying complex cognition such as language production and mind wandering. In a recent study, we have differentiated a Potts associative network into two parts, representing frontal and posterior cortices, to crudely capture with distinct model parameters, including adaptation time scales, salient regional differences observed across mammals. We find that the frontal cortex leads latching in the posterior cortex, determining the sequence of memory states. Here, we analyze the temporal structure of such dynamics.In a homogeneous network, adaptation times set a unique timescale for latching. When they differ sufficiently between the two halves, instead, latching follows the slowly adapting frontal cortex. Interestingly, in the intermediate regime, when the characteristic adaptation times are closer to the ratio of 1:4, we find that latching duration varies within the same sequence, with short and long latches admixed (see Figure). When simple fronto-posterior pattern pairs are then replaced by more complex combinatorial memory structures, we observe that short posterior latches can be nested within long frontal ones, suggesting a mechanism for the spontaneous emergence of long short-term memory (LSTM) functionality at the cortical network level.​Figure: Latching timescales in two halves of a hybrid Potts associative network