The Atkinson-Shiffrin model is ill-defined and does not correctly describe the Murdock free recall data

The Atkinson-Shiffrin (1968) model, the de facto standard model of short term memory cited thousands of times, fits the characteristically bowed free recall curves from Murdock (1962) well. However, it is long overdue to note that it is not a theoretically convincing explanation and that it does not fit all of the experimental relationships in the Murdock data. To obtain a qualitatively correct fit of the bowing I show that four model concepts have to work together. “Long term memory” is needed in the short term memory experiment, conscious or subconscious rehearsal of four items has to take place, this “rehearsal buffer” has to drop items randomly rather than according to a first-in firstout model, and the rehearsal buffer has to be empty before the experiment starts. Beyond the qualitative fit to the bowed recall curves, other relationships in the data are not borne out by the model. First, the “primacy strength”, the ratio of the probability of recall of the first item to the smallest probability of recall of an intermediate item, shows a significant experimental variation with presentation rate but no such variation is predicted by theory. Second, randomly emptying the rehearsal buffer predicts incorrectly that the number of recalled items should be the highest when the first recalled item is the last list item. Third, a simplified Atkinson-Shiffrin model is found to predict exact relationships between the recall probabilities of the initial items which do not seem to be borne out by the Murdock data. Fourth, the theory predicts a discontinuity in the differences between free recall graphs with different presentation rates for early list items which is probably not found in the Murdock data.