Activity in primary sensory areas is robustly driven by signals from their respective sensory inputs. However, both theory and experiment suggest that the same cortices are modulated by context-dependent internally generated activity. Predictive coding theories propose that generative models formed through learning modulate primary sensory responses on a moment-to-moment basis according to how well incoming signals conform to the animal’s predictions. Experimental data has shown that prediction errors can alter neural responses, even in the absence of behavioral relevance. Yet several aspects of how sensory expectation is expressed in primary sensory areas remain unclear. A fundamental question is whether sensory predictions only modulate responses to incoming sensory signals, or if they also reactivate responses based on mnemonic context. Hippocampal populations express past and future events through pre- and replay events, but we are unaware of any work showing such ‘phantasmal’ representations in primary sensory areas. To address this, we chronically recorded neural populations in the primary olfactory cortex during presentation of odor pairs without reinforcement. 12 odors were systematically paired to establish stimulus-specific sensory predictions. In 1/6th of the trials, the second odor was either presented alone (unpredicted) or omitted. We observed changes in the correlation structure and decodablilty of the odor responses during learning. Surprisingly, we observed that phantasmal representation —the ability to decode an odor in absence of actual odor stimuli—developed after just one session for the predicted odor on omission trials. Further, predicted second odor responses were decodable in the population preceding the odor onset. While the source of the mnemonic information—local or from higher areas—is unknown, this data strongly suggests that it provides a complete prediction of the sensory response, as proposed by the theory of predictive coding or in the form of a prior as proposed by Bayesian models.