Memorizing declarative knowledge requires repetition, which can become wearing for learners. In addition, redundant game activities, offering unbalanced challenges in relation to the player's skills, can also lead to a sense of boredom. To reduce this feeling, learning games must provide adapted and varied activities. Automated generation is one way of building such activities. This article proposes a conceptual framework for the design of activity generators for training declarative knowledge in Roguelite games. The framework has been applied in the context of the AdapTABLES project aiming at multiplication tables training.