This paper presents a study of the model of triple BAM by E.Reynaud which is an improved variation of the original BAM model by Kosko. This class of model aims at integrating different sensory inputs in order to memorize a unified and distributed representation. An experimental evaluation of the model is presented that underlines its limitations in terms of noise robustness and learning capacities. A new model is presented in order to overcome those initial limitations by introducing a new online learning algorithm adapted from the PRLAB initial algorithm that improve both noise robustness and learning capacities. Finally, model properties and limitations are considered and discussed within the context of multi-modal integration and brain modeling.