Working memory (WM) allows for the temporary storage of sensory information to guide action. Encoding and maintenance of verbal working memory items involves the phonological loop. We search for a physiological correlate for the replay of items during WM maintenance in auditory cortex and in the medial temporal lobe. In a verbal WM task, 9 epilepsy patients heard strings of letters during encoding and then mentally replayed them during WM maintenance. In each trial, the string consisted of different letter items. While patients performed the task, we recorded from hippocampus, amygdala and enthorinal cortex and extracted single neuron firing (SUA) with automated spike sorting. We further recorded scalp EEG from which we reconstructed the beamforming sources. We computed similarity between encoding and maintainance (EMS) within trials and between trials for both signals. For beamforming sources, we used representation similarity analysis (RSA) on the time-frequency maps. For SUA, we used cosine similarity between firing rate vectors. In SUA, the within-trial EMS exceeded the between-trial EMS (permutation test, p < 0.0005). In EEG beamforming sources, the within-trial EMS built up during encoding and persisted throughout maintenance. The effect was strongest in the auditory cortex of the left hemisphere. Thus, both EEG and SUA indicated that EMS is specific for the WM items. In the EMS we found a physiological correlate for the replay of WM items both in the EEG and in the single neuron firing.