This paper presents a novel model of co-reference knowledge, which is based on the distinction of (i) a model of a common reality, (ii) a model of an agent's opinion about reality, and (iii) a model of agents' opinions if they talk about the same object or not. Thereby it is for the first time possible to describe consistently the evolution of the agent's knowledge and its relative consistency, and to formally study algorithms for managing co-reference knowledge between multiple agents if they have the potential to lead to higher states of consistency, independent from the particular mechanism to recognize co-reference. As an example, a scalable algorithm is presented based on monitoring atomic knowledge increments and an abstract notion of belief ranking. The presented approach has a wide potential to study the formal properties of current methods to find co-reference, and to lead to new methods for the global management of co-reference knowledge.