These nodes would input a list of layers / 3D objects instead of an image. As you fed in a series of layer/object lists into the node over time, it would store them up in the node and use them to compute the feedback. The node would output a grouped layer/object to be passed to Render Layer/Scene to Window. This would contain the most recent layers/objects plus progressively faded-out versions of the previous layers/objects. It would be more computation-heavy than the first option, but would provide correct results for layers / objects beyond the borders.