Four experiments examined the decontextualization of memories, the stage of learning in which memories can be recalled in the absence of episodic memory cues. Face--name pairs were studied with video-recorded environmental contexts in the background, and after 5 practice trials, recall of names associated with faces was tested in the absence of the original video context cues. In Experiments 1, 2, and 3, five retrieval practice trials for a pair occurred either always with the original video context (constant context condition) or with a new context on each trial (varied context condition). Final recall was tested either on the same day (Experiments 3 and 4) or 2 days later (Experiments 1 and 2), and either the recall test context for each face was a new (never seen) video (Experiment 1) or there was no context shown at test (Experiments 2, 3, and 4). In the first 3 experiments, acquisition was better for the constant context condition, but on the final recall test, performance was better for pairs learned under varied context conditions. In Experiment 4, which used multiple study trials rather than multiple retrieval practices during acquisition, no differences were found between the constant and varied context conditions, either for acquisition or for final retention. The results show that acquisition trials given in varied contexts can result in decontextualized memories, but only when acquisition involves retrieval practice, rather than simple restudy trials. These results are consistent with the new theory of disuse, but not with the theory of encoding variability.