This paper reports on a signicifcantly improved version of a system for real-time common sense reasoning. The research is based on the hypothesis that a simple conceptual architecture for memory suffices for a very broad range of behaviors in the common sense world. In particular, we describe a working example of mechanical reasoner that is rather flexible and robust, in that it can tolerate some inconsistencies; can work on goals; can "ruminate" without goals; can forget; can remember; can make assumptions and subsequently detect a conflict between a default conclusion and another assertion (or conclusion), can under suitable conditions decide between them, and can maintain this decision indefinitely until overridden by information to do so.