Current methods for learning Bayesian Networks are mainly batch methods. That is, they are supposed to act in a single step over the complete set of data. We remark the need to develop new approaches that do not require this to happen. Incremental methods do proceed on the supposition that information is fed to the algorithm in a step by step fashion. We propose a formalization for incremental methods, compare it to the most used one in other areas of machine learning and spot several specific peculiarities of Bayesian networks. Present incremental methods are reviewed and criticized in terms of the problems they present for dealing with order effects, and varying sizes of partial data sets. Finally we present BANDOLER a new framework for learning Bayesian Networks incrementally.