Human subjects pointed, without seeing their arm, at visual targets presented in repeated sequences in a frontal plane. Required movement direction could change within the sequence by 0, 45, 90, 135 or 180 degrees. Hand position was recorded contact-free in three dimensions (3D). From the recordings, the pointing errors towards each target were transformed into a Cartesian coordinate system with the x-axis representing the mean direction of all movements towards that target. We then investigated the relationship between successive errors by applying linear regression analysis separately to the three Cartesian error components. For the x-component, we found that successive errors were positively correlated throughout the experiment, which confirms our previous finding that errors in sequential pointing tend to accumulate (Bock and Eckmiller 1986; Bock et al. 1990). Correlation dropped by nearly 50% following a direction change of 90 degrees or more, suggesting that accumulation is reduced but not abolished by large changes in movement direction. The slope of the regression line averaged 0.6, which indicates the existence of a complementary trend towards error correction, contributing about 40% towards motor performance. Changes of movement direction affected slope and correlation in a closely similar way, suggesting that reduced accumulation is paralleled by increased correction. For the y- and z-components, we found that successive errors were positively correlated as well, but were not reduced following even large direction changes. This apparent discrepancy can be resolved by assuming separate neural mechanisms for amplitude and for direction control, differing in their sensitivity to direction changes.(ABSTRACT TRUNCATED AT 250 WORDS)