1. To characterize procedural learning and memory, we devised a behavioral paradigm that allows us to examine the process of learning of new procedures, repeatedly and without serious difficulties for primate subjects. We trained two monkeys to perform a sequential button press task. Upon pressing of a home key, 2 of 16 (4 x 4 matrix) light-emitting diode (LED) buttons (called "set") were illuminated simultaneously, and the monkey had to press them in a predetermined order that he had to find out by trial-and-error. A total of five sets (called "hyperset") was presented in a fixed order for completion of a trial; an error at any set aborted the trial. A given hyperset was repeated as a block of experiment until 20 successful trials were performed. Monkeys PI and BO experienced 313 and 92 hypersets, respectively. Most of these hypersets were experienced only once (1 block of experiment); the others (28 hypersets for monkey PI and 14 hypersets for monkey BO) were chosen for extensive practice. 2. The learning, indicated as the decrease in the number of trials to criterion and the decrease in the performance time, proceeded at three levels: 1) short-term and sequence-selective learning that occurred by repeating a particular hyperset during a block of experiment; our monkeys learned, to some degree, to perform a new hyperset within a short period (< 5 min); 2) long-term and sequence-selective learning that took place for each hyperset across days; by daily practice, they further improved their skills for performing the particular hyperset; and 3) long-term and sequence-unselective learning that was indicated by the improvement of performance for new hypersets; the monkeys were required to learn many hypersets, each just once (a block of trials), in which they performed gradually better with more experiences in the 2 x 5 task. 3. To examine whether the memory was retained for a long period, we had the monkey learn 12 hypersets sufficiently, then we stopped the training and retested them after 1 or 6 mo. After the 1-mo interruption the performance was significantly better than that for new hypersets. After the 6-mo interruption the performance was not different from new hypersets in terms of the number of trials but was significantly better than new hypersets in terms of the performance time. The results suggest that motor memory (measured by performance time) can be retained longer than procedural memory (measured by the number of trials).