Predictions of the climate of the next few decades are of increasing importance for science, businesses and politics. Most existing predictions are based entirely on past and projected changes in radiative forcing, and make no use of information about the current climate state. However, recent work at the UK Hadley Centre (Smith et al, submitted to Science) has shown that assimilation of upper ocean observations can increase the skill of decadal predictions. In this study we are investigating the mechanisms that give rise to increased skill, with a specific focus on the role of North Atlantic heat content. Observations suggest the North Atlantic has exhibited large decadal variability in recent times, and the implications of this variabilty for predictions have yet to be established. We make use of the HadCM3 climate model, including a representation of anthropogenic and natural aerosols. We are carrying out ensemble hindcasts from model states that correspond to a common time (based on a common history of changing external forcings, and a common start date, e.g. 1990), but differ with respect to the state of the North Atlantic Ocean. By comparing the evolution of climate variables in the two ensembles we are able to identify and quantify the impact of the Atlantic Ocean state, and hence the potential to improve predictions through the greater and better use of ocean observations.