A reduced-pitch-pattern model for melodic processing is proposed. The model assumes that experienced listeners divide a melody into segments, derive a reduced pitch pattern from each segment, and then try to match each pattern to one of the prototypes stored in long-term memory. As a result, the melody is memorized accurately and quickly. This model and two competing models (the contour model and the harmonic progression model) make different predictions concerning what aspects of a melody would be preserved and what types of error in recall would appear. Recall data of a tonal melody of 12 measures in length were used to examine these predictions. The data were gathered from eight college music majors. Analysis of erroneous reproductions showed that reduced pitch patterns and harmonic progressions were preserved well and that errors predicted by the reduced-pitch-pattern model occurred more often than those predicted by the two competing models. It is concluded that the reduced-pitch-pattern model is the most tenable of the three.