Healthy adult humans performed elbow movements in a horizontal plane under a small external extending torque (2.1-3.3 Nm). Test movements (TMs) consisted of slow ramp-and-hold flexions in the absence of visual feedback, with the target joint angle to be remembered from a preceding conditioning movement (CM). The CM was produced by matching two beams on the monitor screen: (1) command representing the target position (a straight line); and (2) a signal from the sensor of the elbow joint angle. Two kinds of CM were applied, which had the same target position (50 degrees in most experiments) but differed in initial positions: (1) fully extended joint (0 degrees, P1 CMs); (2) flexed joint (100 degrees, P2 CMs). In a group of 25 subjects, the target in TMs was usually overshot, with the position errors depending on the CMs: 2.7 +/- 0.6 degree (mean +/- SEM) for P1 CMs, and 10.9 +/- 0.7 degree (P < 0.001) for P2 CMs. Vibration of the elbow flexors substantially diminished the difference between the position errors, amounting to--0.31 +/- 0.5 degree and 2.33 +/- 0.6 degrees, respectively. It is suggested that the observed position errors resulted from after-effects in the activity of muscle spindles in agonist and antagonist muscles, but influence of differences in dynamic components of the afferent signals during oppositely directed approaches to the target cannot be excluded.