A fundamental challenge in neuroscience is to understand the mechanisms by which multi-component actions are represented and sequenced for production. We addressed this challenge with a movement imitation task in which subjects viewed the quasi-random, 2D movements of a disc and then used a stylus to reproduce the remembered trajectory. The stimulus disc moved along straight segments, which differed sufficiently from one another that it was possible to trace individual segments' fate in the resulting movement imitation. A biologically-based segmentation algorithm decomposed each imitation into segments whose directions could be compared to those of homologous segments in the model. As the number of linked segments in a stimulus model grew from three to seven, imitation became less accurate, with segments more likely to be deleted, particularly from a model's final stages. When fidelity of imitation was assessed segment by segment, the resulting serial position curves showed a strong primacy effect and a moderate recency effect. Analysis of pairwise transposition errors revealed a striking preponderance of exchanges between adjacent segments, which, along with the serial position effects, supports a competitive queuing model of sequencing. In analogy to results with verbal serial recall, repetition of one directed segment in the model reduced imitation quality. Results with longer stimulus models suggest that the segment-by-segment imitation generator may be supplemented in the final stages of imitation by an error-signal driven overlay that produces a late-course, real-time correction. Results are related to neural mechanisms that are known to support sequential motor behavior and working memory.