Current OO frameworks provided with MAS development toolkits provide core abstractions to implement the agent behavior, by using the typical OO specialisation mechanisms. However, these OO designs do not provide proper abstractions to modularize other extra-functional concerns (e.g. learning property), which are normally intermingled with the agent functionality (tangled code problem), and spread over different classes or components (crosscutting concerns problem). This means that the reusability of the agent architectural components is drastically reduced, so agents are difficult to maintain, extend or adapt. Aspect-oriented technologies overcome these problems by modeling such concerns as aspects. This work proposes to separate and modularize the learning of software agents following the aspect-oriented solution of the Malaca model. By decoupling the agent functional behavior from the protocol that carries out the learning activities; the development, adaptation and evolution of intelligent agents is substantially improved.