The motor system must be able to adapt quickly to load changes. In addition, it is known that also long-term learning exists in the motor system, which, for example, allows us to learn difficult motor skills like playing the piano or excelling in sports. It is generally agreed that many times motor learning occurs through the generation of forward models for an originally existing control loop. Hence, early on, the system cumbersomely executes a motor action under tight closed-loop feedback control, whereas later, after some successful learning, a forward model has been generated and the system executes a much improved motion sequence "without thinking". In this study we show that it is possible to combine temporal sequence learning with recruitment to learn a forward model of a reflex and to execute it with the momentarily required strength, which depends on the load. To this end we employ a learning rule, which we had recently introduced (Porr and Woergoetter Neural Comp. 15, 831-864, 2003) and which is similar to spike timing-dependent plasticity. Using this rule, at a one-arm joint we correlate an earlier disturbance loading force with the measured position change that this force induces. Without learning, the position change is originally compensated with some delay by a reflex like feedback reaction. After learning the system reacts immediately to the loading pulse and produces the required counterforce to keep the arm position stable. This works for constant loads. It is, however, possible to add a simple recruitment mechanism to this and now the system can learn compensating also different loads without delay. From an older study (Porr et al, Neural Comp. 15,865-884,2003) a mathematical proof exists that this type of learning creates a forward model of the initially existing feedback loop. Hence, here we were able to show that we can employ learning and recruitment in a simple motor-control structure to learn anticipatory load compensation by feed-forward compensation.