A model for memory scanning is proposed in which the encoded representation of a probe is compared in parallel with encoded representations of each item in the positive set. The within item matches are serial feature by feature comparisons that terminate when either a positive or negative criterion is reached. This model is shown to predict the results of a probe similarity experiment. The serial location of a similarity within an item affects negative reaction times, but the number of items in the positive set to which the probe is similar has no main effect. The model is also shown to yield predictions consonant with existing data on the relation between reaction times and set size and speed-accuracy trade offs. In a memory scanning recognition task a subject must decide as rapidly as possible whether or not a presented test item was included in a previously presented set of items. Typically, a subject is asked to remember a list, s, of digits, letters or words presented visually or auditorily. He may or may not be allowed to rehearse the list. On a test trial he is presented with a test item, d, and asked to decide whether it was in the positive set, s. Usually, reaction time is the dependent variable of interest, and performance is designed to be nearly error free. Sternberg's (1966, 1969) well known exhaustive scanning model for explaining recognition times in memory scanning was based on three major findings from his early experiments. First, recognition times increase linearly as a function of positive set size. Second, the slope of the positive and negative reaction times plotted against set size are about equal. Third, positive recognition times are independent of the serial position of the