One of the characteristics of cursor movement controlled via a brain machine interface is a trade-off between the ability to move rapidly between targets and the ability to hold the cursor steadily within a target. We propose to address this limitation by classifying independent movement and posture states, and using neural decoders with optimum dynamical properties for each state. This paper investigates two methods of classifying the state of a limb based on the offline analysis of neural discharge. We also tested the performance of state-dependent decoders that either apply additional smoothing during the posture state or consist of separate filters trained explicitly on data from the different movement states. This work suggests that a state-dependent decoder may provide significantly improved BMI performance.