The mismatch negativity (MMN) component of the auditory event-related potential can be used to study automatic perceptual inference. This study was designed to explore “Conditional Inference” in the MMN system—the capacity of the auditory system to use current input to switch between inference models in memory. We presented a “Random” sequence comprising a standard repeating sound occasionally interrupted by a change in frequency, duration or intensity (louder or softer). We also presented the same sequence with a conditional linkage between deviants—that is, frequency and duration deviants always followed an intensity change. We explored whether the auditory system could use intensity deviance to change the inference from “expect the standard to repeat” to “expect a frequency or duration violation” and quantified this as the proportion decline in the MMN elicited to duration and frequency deviant sounds in the linked versus random sequence. We report three main outcomes on a sample of 25 healthy young adults: (1) there was a significant conditional inference effect (a reduction in MMN amplitude in linked versus random sequences) to duration but not frequency deviants; (2) larger simple MMN and larger conditional inference effect on duration MMN were correlated with higher Digit Span; and (3) the conditional inference effect but not simple MMN to duration deviants was strongly correlated with working memory ability (rs = 0.78, p < 0.001). The results are discussed with respect to the differential cognitive demands of simple MMN and conditional inference, and the possible involvement of prefrontal cortex in implementing conditional inference in the MMN system.