Agents that learn from experience can profit immensely from memorizing what they have done, why, how, and what happened. For autonomous robots performing complex manipulation tasks, these memories include low level data, such as perceptual snapshots of relevant scenes that influenced decision making, detailed complex motions the robot performed, and effects of these motions. They also include high level representations of the intended actions and the belief-dependent descisions that led to the chosen course of action. In this paper, we present CRAMm, a memory management system that records very comprehensive and informative memories without slowing down the operation of the robot. CRAMm offers a query interface that lets the robot retrieve the kinds of information stated above. This is done using a first-order logical language that provides predicates concerning the beliefs and intentions of the robot, its physical state, perceptual information, and action effects, as well as their relations at different levels of abstraction.