Attempting to guess an answer to a memory question has repeatedly been shown to benefit memory for the answer compared to merely reading what the answer is, even when the guess is incorrect. In this study, we investigate 2 potential explanations for this effect in a single experimental procedure. According to the semantic explanation, the benefits of guessing require a clear semantic relationship between the cue, the guess, and the target, and these benefits arise at the stage of guessing. The attentional explanation places the locus of the effect at the stage of feedback presentation and ignores the issue of semantic relatedness. To disentangle the 2 mechanisms, we used homograph cues with at least 2 different meanings (e.g., "arms") and asked participants to either study an intact cue-target pair or guess a word related to each cue before being presented with the target. This allowed us to compare memory performance on trials in which participants' guesses tapped the same meaning of the cue as the later presented target (e.g., a guess "legs" for a pair "arms-hug"), versus a different meaning (e.g., "weapons"). In 4 experiments, we demonstrated that both the semantic and the attentional mechanism operate in the guessing task, but their roles are different: Semantic relatedness supports memory for cue-to-target associations, whereas increased attention to feedback benefits memory for targets alone. We discuss these findings in the context of educational utility of errorful learning.