Multi-limb motor skills, such as swimming and rowing, often involve isolated practice of each limb (unimanual) followed by practice with both limbs together (bimanual). We recently demonstrated that learning a novel load during unimanual reaching is partially, but not completely transferred to the same limb during bimanual reaching (and vice versa), learning can remain hidden and only revealed by the original context, and the ability to learn two conflicting force fields if each was separately associated with unimanual and bimanual reaching (Nozaki et al. 2006). The purpose of the present article is to develop a formal state-space model to conceptualize and interpret these complex experimental results. The model contains three separate compartments for learning, a unimanual-specific, a bimanual-specific, and an overlapping compartment, and the internal state of each compartment is updated context-dependently according to motor errors. The model was able to capture all major aspects of motor learning across these two behaviours and predict further complexities during washout trials when bimanual and unimanual trials are interleaved. We propose that partial, but not complete transfer of motor learning is due to a corresponding partial overlap in neural control processes across these behaviours, and is a general feature of different classes of voluntary motor behaviour, such as postural control, point-to-point reaching, manual tracking and oscillatory movements.