We propose that memory confidence is based on the processes and products of the just-completed memory task, along with the participants' metamemory beliefs about the relation of these processes and products to memory accuracy. We tested this metamemory approach to confidence by having participants carry out a simple recognition memory task with deceptive and nondeceptive items. The deceptive items were sentences that contained a possible synonym substitution, thus allowing errors based on gist memory. For nondeceptive items, high confidence was associated with high accuracy. For deceptive items the relationship reversed; high confidence was associated with low accuracy. A memory process questionnaire was developed that provided more differentiated phenomenal reports than the traditional know/remember distinction. For nondeceptive items, metamemory beliefs tended to be valid indicators of accuracy, but for deceptive items involving unconscious reconstructive memory processes, they tended to be invalid indicators of memory accuracy. The overall results lend strong support for our metamemory approach to memory confidence.