Several theories assume that practice (a) results in restructuring of component processes and (b) reduces demand on working memory. Eight subjects practiced judgments about digital logic gates for over 8,000 trials. At two practice levels, subjects made judgments while retaining short-term memory loads irrelevant to the judgments, relevant but not accessed, or accessed to make the judgments. Four phenomena together provide constraints for theory: First, performance declined in moving from blocked practice to randomized practice. Second, gate and judgment type strongly affected latency. Third, these effects declined but did not disappear with practice. Fourth, the cost of accessing information in working memory remained substantial. These results are interpreted as reflecting a serial process with constant structure, while component processes become faster. The results challenge theories assuming that all learning results from restructuring or that restructuring is an automatic consequence of practice, and they support a distributed view of working memory.