Stimulus features that occur in action episodes are integrated with response features into event files, and event files are retrieved when features are repeated. However, the strength of retrieval depends on the extent to which a particular feature is actually repeated. Since not only perceptual but also conceptual features are assumed to be bound, the boundaries between feature representations may not only be fluid but also modulable. In three experiments (cumulative n = 217), we investigated whether internal manipulation of feature boundaries is possible. By adding additional counting tasks to a stimulus-response binding task, we either merged or separated feature categories, causing a significant difference in S-R binding effects: Merged feature categories produced weaker S-R binding effects compared to separated feature categories. The results suggest that merged features were actively made to be processed as more similar to each other. We interpret our data within the broader and old question of what a feature actually is, and suggest that feature boundaries are task-dependent.