During perceptual decision-making, neural circuits can accumulate decision evidence and maintain the choice in short-term memory until the motor response. Previous modelling work has proposed that recurrent circuit dynamics underlie these two processes and that N-methyl-D-aspartate receptors (NMDARs) are necessary for proper network function. Electrophysiological evidence indicates that circuits in the frontal cortex may carry out these computations but the role of NMDAR has not been systematically tested. We developed an auditory two-alternative delayed-choice (2ADC) task for head-fixed mice involving the accumulation of time-varying auditory evidence as well as its maintenance during a delay mnemonic period. Choice accuracy increased with stimulus evidence S (i.e. displayed psychometric curves) and decreased with imposed delay length D, a signature of forgetting. Mice also showed a tendency to repeat their previous choices regardless of S or D. We fitted this behaviour using a modified double-well model and compared it with a two-state Hidden Markov model (HHM), in which mice were either in an engaged state where choices were guided by the stimuli or in a disengaged state in which mice did repeating lapses. Model comparison favoured the HMM as a better model. Systemic administration of an NMDAR antagonist decreased choice accuracy without impairing evidence accumulation or memory maintenance. Instead, it decreased the probability of being in the engaged state. Thus, our results show that NMDAR blockade may affect the cognitive control necessary for mice to accumulate and maintain relevant task information rather than impacting the specific dynamics of those processes.