Many recent studies have shown that memory for correct answers is enhanced when an error is committed and then corrected, as compared to when the correct answer is provided without intervening error commission. The fact that the kind of errors that produced such a benefit, in past research, were those that were semantically related to the correct answer suggested that the effect may occur because the error provides a semantic stepping stone to the correct answer: the Semantic Mediation hypothesis. This hypothesis seems at odds with the finding that amnesicsgenerate answers, again including those studied by Tulving and his colleagues--who purportedly have spared semantic/implicit memory-experience enormous difficulties when they commit errors. Accordingly, the present experiments investigated whether the error-generation benefit seen in typicals was attributable Semantic Mediation or to Episodic Recollection. In Experiment 1, we used polysemous materials to create Congruent (e.g., wrist-palm) and Incongruent (e.g., tree-palm) cues for target words (e.g., HAND). In the Congruent condition, participants generated errors that were semantically related to the target (e.g., finger), and which could have provided a semantic mediator. In the Incongruent condition they generated errors that were "unrelated" to the target (e.g., coconut), and which, therefore, should not have provided a semantic mediator. The Congruent and Incongruent conditions both produced an error-generation benefit--contradicting the Semantic Mediation hypothesis. Experiment 2 showed that the error-generation benefit only occurred when the original error was also recollected on the final memory test. Indeed, in the Incongruent condition, when the error was not, itself, recalled, error generation resulted in a deficit in memory for the correct response. These results point to episodic/explicit, rather than semantic/implicit memory, as the locus of the 'learning from errors' benefits.