Two experiments examined the temporal specificity of learning in operant serial feature-positive discriminations (feature-->target+/target-). Test performance was better when the target cues were presented at their customary times after the features than when they were presented at earlier or later times. When features trained with one feature-target interval were combined with targets trained with another interval, performance was best when the test interval was the same as the interval associated with the feature, suggesting that the temporal information was coded with the feature cue. Finally, the temporal specificity of the responding controlled by occasion setters was unaffected by feature extinction. Implications for the nature of learning in occasion setting are discussed.