In the task-switching paradigm, response repetitions (RR) usually yield performance benefits compared to response switches, but only when the task also repeats. When the task switches, RR benefits vanish or even reverse into costs. This interaction between task and response repetition versus switch is called the RR effect. Several theoretical accounts have been proposed for explaining the RR effect. Here, we tested a prediction derived from episodic-retrieval accounts, namely that RR benefits in task repetitions occur because repeating the task retrieves the task-response binding formed in the previous trial. To this end, we considered the probability that participants erroneously repeated the previous-trial response in response-switch trials (i.e., response-retrieval errors). Thus, our tasks employed three response alternatives in order to discriminate between response-retrieval errors and other errors in response-switch trials. Across two task-switching experiments (N= 46 and N=107), results showed that, in response-switch trials, response-retrieval errors were more likely in task repetitions than switches, supporting the notion that the previous response is retrieved by the repeating task, despite being wrong. This finding cannot be easily accommodated by the competing theoretical accounts (e.g., the response-inhibition or the associative-learning account). Thus, the present study indicates task-response binding as an important mechanism behind the RR benefits in task repetitions, since a repeating task seems to retrieve the previous response, which is correct in response repetitions but wrong in response switches.