Reaction time (RT), that is the time it takes for an action plan to be executed, has been the interested of numerous studies. Of particular focus has been the RT in the context of perceptual ambiguities, when evidence towards taking a certain action must be integrated over time in order to execute the correct action. The stochastic drift diffusion model explains how such integration leads to the observed variability in RT, but what explains RT variability in a task where there is no ambiguity in what action will be rewarded?We trained recurrent neural network models to perform a delayed memory saccade task, and found that these networks replicated several key aspect from neural data of non-human primates (NHP) trained on the same task. In particular, we found units that encoded information about the remembered target, as well as about the upcoming action right before and during action execution. Such units have also been found in the prefrontal areas of NHPs, and are thought to play a key role in the transition from a memory maintenance state to an action execution state.By analysing the connectivity matrix of the trained models, we were able to trace the flow of information through the network as the models were presented with different target locations to memorise and later report. This valuable insight could shed light on how *in vivo* neural circuits carry out similar information transformations, and therefore aid in the understanding of how higher cognitive processes are implemented in the brain.