Three experiments were run in which subjects first memorized examples of input-output pairs and then generated the outputs for a series of new inputs by analogy to the original examples. Subjects first performed these mappings by explicit analogy to an example but with practice they learned to make these input-output mappings directly without reference to the examples. Subjects sped up as a power function of practice over a day (Exp 1) or days (Exps. 2 and 3). In the multi-day experiments, each day required a few trials of practice before subjects returned to the speed of the former day. Subjects developed a directional asymmetry such that they were slower to calculate the input from the output than the output from the input (whereas initially they had not been). Subjects showed similar speed up in their ability to recall the original examples but did not show the same directional asymmetry. Initially, there was some transfer from practicing the procedure to recalling the examples but this diminished over days. The implications of this research are discussed for the relationship between procedural and declarative knowledge in the ACT theory (Anderson, 1983, 1993).