Framing metacognitive judgments of learning (JOLs) in terms of the likelihood of forgetting rather than remembering consistently yields a counterintuitive outcome: The mean of participants' forget-framed JOLs is often higher (after reverse-scoring) than the mean of their remember-framed JOLs, suggesting greater confidence in memory. In the present experiments, we tested 2 competing explanations for this pattern of results. The optimistic-anchoring hypothesis suggests that forget-framed JOLs are associated with greater optimism about memory than are remember-framed JOLs, which leads to their greater magnitude. The differential-scaling hypothesis suggests that forget-framed JOLs and remember-framed JOLs will often be distributed differently across the JOL scale, resulting in means that also often differ. Participants in 3 experiments studied simple memory materials and made JOLs predicting their memory performance for those items. They made their JOLs in terms of either the likelihood of remembering or forgetting. In contrast to the optimistic-anchoring hypothesis, the mean of participants' forget-framed JOLs was unaffected by information concerning the supposed difficulty of the task (Experiment 1), was lower than for remember-framed JOLs in a task selected to evoke high JOLs (Experiment 2), and demonstrated equivalent confidence in memory when participants were restricted to a yes--no binary response (Experiment 3). In support of the differential-scaling hypothesis, participants' forget-framed JOLs were consistently symmetrically distributed across the JOL scale, resulting in a mean at the center of the judgment scale that was often higher than that for remember-framed JOLs. Framing therefore affects how participants scale their JOLs, not their confidence in their memory.