While most of the current work has been focused on developing adaptive techniques to respond to human-initiated inputs (what behaviour to perform), very few of them have explored how to proactively initiate an interaction (when to perform a given behaviour). The selection of the proper action, its timing and confidence are essential features for the success of proactive behaviour, especially in collaborative and assistive contexts. In this work, we present the initial phase towards the deployment of a robotic system that will be capable of learning what, when, and with what confidence to provide assistance to users playing a sequential memory game.