Errorless learning has been shown to be very successful in the rehabilitation of memory problems particularly in patients with severe forms of memory impairment. Much of this research has focused on testing knowledge of specific details studied, ignoring any additional, higher-level knowledge that patients may have acquired during the learning process. Hence, it is pertinent to ask whether errorless learning is equally successful in the acquisition of high and low-level knowledge. In this paper, we present results of several studies comparing the effectiveness of errorless and standard trial-and-error methods in acquisition of high and low-level knowledge in people diagnosed with dementia and non-impaired controls. In Study 1, participants were asked to learn novel face–name–occupation associations; and knowledge across a range of levels, from very general (i.e., high-level) to very specific (i.e., low-level), was examined. For patients with probable Alzheimer's disease and controls there was evidence of increased benefit from errorless training in general, but the technique was most beneficial for patients attempting to retrieve specific detail. Study 2 was conducted to address the problem raised by the failure in Study 1 to manipulate learning condition at our highest knowledge level. This novel manipulation was successful, but neither of the patients received the standard benefit from errorless training. Study 3, involving a small group of dementia patients with mixed diagnoses, was conducted to replicate findings from Study 1. Results from the group analysis confirmed that the benefit obtained from errorless learning increased as a function of knowledge specificity, but again several patients failed to show a consistent effect of learning condition. Implications for use of the errorless technique are discussed.