The purpose of this study was to characterize the nature and structure of procedural memory. We have previously studied the process of learning sequential behavioral procedures using monkeys. The monkey's task was to press five consecutive pairs of buttons (indicated by illumination) in the correct order for every pair, which he had to find by trial-and-error in a block of trials. The whole sequence was called a "hyperset"; each pair was called a "set". We first examined whether monkeys learned to perform a hyperset as a single sequence or learned the order of button-presses individually for each set. To answer this question, we generated hypersets that were the same as the hypersets that had been extensively learned except that the order of the sets was reversed. The performance of these "reversed hypersets" was much worse than the performance of the original learned hypersets and was similar to the performance of new hypersets, as regards both the number of errors and the performance time. The result suggests that monkeys learned a hyperset as a sequence. To examine whether the learned performance was specific to the hand used for practice, we had monkeys use the same hand throughout the long-term practice of each hyperset, and then tested the opposite hand. The performance using the opposite hand was worse than the performance using the trained hand, but was better than the performance for new hypersets. This indicates that the memory for the sequential procedure is only partially accessible to the hand that was not used for the practice.