Working memory (WM) is central for cognition and is impaired in many brain disorders including those hypothetically mediated by NMDA receptor (NMDAR) hypofunction. Evidence suggests that network attractor states underlie WM maintenance and that failures are mostly caused by fluctuation-driven transitions, but direct evidence for this is still lacking. To investigate what makes WM fail we used a simple two-alternative delayed response task in which mice listen to a lateralized auditory stimulus and, after a variable delay (duration D=0-10 s), they have to lick the associated lateral port. Mice accuracy decreased with delay showing that there were forgetting errors. They also showed a repeating bias, i.e. a tendency to repeat the previous choice, which was however independent of delay. Inactivation of NMDAR caused a decrease in accuracy and an increase of the repeating bias, but critically did not affect the forgetting rate, i.e. the decay of accuracy with delay. We recapitulate these findings using a hidden Markov Model that switches between (1) a WM state describing a stimulus-based strategy that requires memory maintenance, and (2) a history-based state (HB) which elicits lapse responses determined by previous choices. A shift in the transition probabilities towards the HB-state reproduces the effect of NMDAR blockade. Electrophysiological recordings in the anterolateral motor cortex (ALM) supported our model by showing that the encoding of the stimulus differs between inferred HB-state and WM-state trials: neurons showed similar encoding during the stimulus presentation, but only in the WM-state the encoding persisted during the delay period. Our results show that task performance is heavily limited by lapse epochs in which subjects stop using WM and suggest that deficits caused by NMDAR hypofunction could be related to an increased sensitivity to the cognitive effort associated with WM rather than by a decrease in its stability.