Judgment of learning (JOL) plays a pivotal role in self-regulated learning. Although the JOLs are in general accurate, important deviations from memory performance are often reported, especially when the JOLs are made immediately after learning. Nevertheless, existing studies have not clearly dissociated the neural processes underlying subjective JOL and objective memory. In the present study, participants were asked to study a list of words that would be tested one day later. Immediately after learning, participants predicted how likely they would remember that item. Critically, the JOL was performed on only half of the studied items to avoid its contamination on subsequent memory. We found that during encoding, compared to items later judged as will be forgotten, those judged as will be remembered showed stronger activities in the default-mode network, including the ventromedial prefrontal cortex (PFC) and posterior cingulate cortex, as well as weaker functional connectivity between the left dorsolateral PFC and the visual cortex. The exact opposite pattern was found when comparing items that were actually remembered with that were later forgotten. These important neural dissociations between JOL and memory performance shed light on the neural mechanisms of human metamemory bias.