We assume a system in which a robot maintains its information about the world and itself primarily as a collection of sentences in a mathematical logical language. There will be other data structures where they is more compact or computationally easier to process, but they will be used by programs whose results become stored as sentences. The robot decides what to do by logical reasoning, not only by deduction using rules of inference but also by nonmonotonic reasoning.
We do not attempt a full formalization of the rules that determine the effects of mental actions and other events in this paper. The main reason is that we are revising our theory of events to handle concurrent events in a more modular way. There is something of this in the draft [McCarthy, 1995a].
Robot consciousness involves including among its sentences some about
the robot itself and about subsets of the collection of sentences
itself, e.g. the sentences that were in consciousness just previous to
the introspection, or at some previous time, or the sentences about
a particular subject. 
We say subsets in order to avoid self-reference as much as possible. References to the totality of the robot's beliefs can usually be replaced by references to the totality of its beliefs up to the present moment.