Using two bias reduced forcing data sets we have simulated the initial soil moisture state for the North American continent (1985--1993). Differences between simulations were shown to persist over regions with the greatest soil memory, and are not strongly associated to patterns of difference in forcing fields. Therefore, processes that contribute to soil memory will also limit our ability to accurately estimate its initial state. Moreover, the regions associated with high soil memory may not be directly observable with present and near future sensors due to high vegetation water contents. Thus, research in these regions should be directed toward minimizing forcing variance and extension of observation networks and technology to increase the ability for monitoring soil moisture.