Persistent activity has commonly been considered to be a hallmark of working memory (WM). Recent evidence indicates that neuronal discharges in the medial temporal lobe (MTL) are compatible with WM neural patterns observed in cortical areas. However, the characterization of this activity rarely consists of measurements other than firing rates of single neurons. Moreover, a varied repertoire of firing dynamics has been reported in the MTL regions, which motivate the more detailed examination of the relationships between WM processes and discharge patterns undertaken here. Specifically, we investigate at different resolution levels firing irregularities in electrode recordings from the hippocampus, amygdala, and the entorhinal cortex of epileptic patients during a WM task. We show that some types of (ir)regularities predict response times of the patients depending on the trial periods under consideration. Prominent burst activity at the population level is observed in the amygdala and entorhinal cortex during memory retrieval. In general, regular and bursty neurons contribute to the decoding of the memory load, yet they display important differences across the three anatomical areas. Our results suggest that non-random (non-Poisson) patterns are relevant for WM, which calls for the development and use of statistics complementary to mere spike counts.