A cognitive hierarchical storage-management system receives feedback describing users' satisfaction with the way that one or more prior data-access requests were serviced. The system uses this feedback to associate each previously requested data element's metadata and storage tier with a level of user satisfaction, and to optimize user satisfaction when the system is trained. As feedback continues to be received, the system uses machine-learning methods to identify how closely specific metadata patterns correlate with certain levels of user satisfaction and with certain storage tiers. The system then uses the resulting associations when determining whether to migrate data associated with a particular metadata pattern to a different tier. Data elements may be migrated between different tiers when two metadata sets share metadata values. A user's degree of satisfaction may be encoded as a metadata element that may be used to train a neural network of a machine-learning module. If detecting that two metadata sets share metadata values, the system determines whether to migrate data elements to different tiers.