A repeated finding in the literature of face recognition is that expressive faces are remembered better than neutral faces. However, a better facial-identity recognition may come at a cost of a reduced precision with which the pictorial facial features, irrelevant for identity recognition, are represented in memory. By means of a continuous-report task, we tested this hypothesis by measuring the memory precision of expressive and neutral faces. Commensurable face-identity and facial-expressions variations were generated with the method of Fechnerian scaling. The results confirm our hypothesis, but only under conditions of high memory load. We interpret the present findings as due to the effects of the categorical processes required for facial-identity recognition.