Of the many connotations of 'rational' we focus on 'optimality': Given the constraints imposed by what has been stored in memory, is retrieval from memory optimal or close to optimal, and is retrieval in different tasks constraints imposed by what has been stored in memory, is retrieval from memory optimal or close to optimal, and is retrieval in different tasks equally optimal or equally far from optimal? Such questions can be addressed in the context of models that are based on probabilities of matching probe cues to memory traces, and we couch our discussion in the context of one such model, REM. Optimality is discussed first for explicit single-item yes-no recognition. We show that relative optimality is less when extensions are made to tasks in which the probe consists of more than one item, such as paired recognition, and associative recognition. We argue that retrieval is probably less optimal in recall than recognition. Extensions to generic and implicit memory are considered briefly.