Previous works in Brain-Machine Interfaces (BMI) have mostly used a single Kalman filter decoder for deriving continuous kinematics in the complete execution of behavioral tasks. A linear dynamical system may not be able to generalize the sequence whose dynamics changes over time. Examples of such data include human motion such as walking, running, and dancing each of which exhibit complex constantly evolving dynamics. Switching linear dynamical systems (S-LDSs) are powerful models capable of describing a physical process governed by state equations that switch from time to time. The present work demonstrates the motion-state-dependent adaptive decoding of hand and arm kinematics in four different behavioral tasks. Single-unit neural activities were recorded from cortical ensembles in the ventral and dorsal premotor (PMv and PMd) areas of a trained rhesus monkey during four different reach-to-grasp tasks. We constructed S-LDSs for decoding of continuous hand and arm kinematics based on different epochs of the experiments, namely, baseline, pre-movement planning, movement, and final fixation. Average decoding accuracies as high as 89.9%, 88.6%, 89.8%, 89.4%, were achieved for motion-state-dependent decoding across four different behavioral tasks, respectively (p<0.05); these results are higher than previous works using a single Kalman filter (accuracy: 0.83). These results demonstrate that the adaptive decoding approach, or motion-state-dependent decoding, may have a larger descriptive capability than the decoding approach using a single decoder. This is a critical step towards the development of a BMI for adaptive neural control of a clinically viable prosthesis.