Pointing to a remembered visual target involves the transformation of binocular visual information into an appropriate motor output. Errors generated during pointing tasks may indicate the reference frames used by the CNS for the transformation and storage of the target position. Previous studies have proposed eye-, shoulder-, or hand-centered reference frames for various pointing tasks, depending on visual conditions. We asked subjects to perform pointing movements to remembered three-dimensional targets after a fixed memory delay. Pointing movements were executed under dim lighting conditions, allowing vision of the fingertip against a uniform black background. Subjects performed repeated movements to targets distributed uniformly within a small (radius 25 mm) workspace volume. In separate blocks of trials, subjects pointed to different workspace regions that varied in terms of distance and direction from the head and shoulder. Additional blocks were performed that differed in terms of starting position, effector hand, head rotation, and memory delay duration. Final pointing positions were quantified in terms of the constant and variable errors in three dimensions. The orientation of these errors was examined as a function of workspace location to identify the underlying reference frames. Subjects produced anisotropic patterns of variable error, with greater variability for endpoint distances from the body. The major axes of the variable-error tolerance ellipsoids pointed toward the eyes of the subject, independent of workspace region, effector hand (left or right), initial hand position, and head rotations. Constant errors were less consistent across subjects, but also tended to point toward the head and body. Both overshoots and undershoots of the target position were observed. Increasing the duration of the memory delay period increased the size but did not alter the orientation of the variable-error ellipsoids. Variability of the endpoint positions increased equally in all three Cartesian directions as the memory delay increased from 0.5 to 8.0 s. The anisotropy of variable errors indicates a viewer-centered reference frame for pointing to remembered visual targets with vision of the finger. The anisotropy of pointing variability stems from variability in egocentric binocular cues as opposed to reliance on allocentric visual references or to specific approximations in the sensorimotor transformation. Nevertheless, observed increases in variability with longer memory delays indicate that the short-term storage of the target position does not simply mirror the retinal and ocular sensory signals of the visually acquired target location. Thus spatial memory is carried out in an internal representation that is viewer-centered but that may be isotropic with respect to Cartesian space.