Although eyewitness testimony is often unreliable, its quality is not uniformly poor. Previous research indicates that the most accurate type of testimony involves the unstructured recall of actions. The present study replicates the finding that confabulation, particularly the complete fabrication of nonexistent actions, is rare. Our results also extend earlier work by revealing that the few errors that did occur were spread evenly across subjects—with most subjects making 0 or 1 error and with no subjects making more than 3 errors. Furthermore, action memory errors form a heterogeneous set—there were few commonly recurring or predictable errors. Our attempts to increase error frequency by the use of suggestive implications and rapid action sequences were unsuccessful. Practical aspects of evaluating action memory testimony are discussed, with special reference to cases involving multiple eyewitnesses.