Agent-based distributed problem solving environments decompose a given problem into a set of sub problems which can be processed in parallel and independently by autonomous and mobile agents at each computation node. Such an autonomous agent primarily makes use of local information which is provided at the respective computation node. This kind of information is characterized by its potential incompleteness and inconsistency with regard to the overall distributed system state which is due to the lack of any centralized coordination facility. In this paper, we introduce the use of long-term knowledge repositories for autonomous agents without sacrifying the autonomy of the agents and without introducing any central management facility. We compare our approach to an agent-enabled distributed SAT prover which makes only use of local system state information. In that problem solving application a given search tree is distributed dynamically by autonomous mobile agents implemented in pure Java. To demonstrate the profit of using knowledge repositories in general, we integrated our agent system Okeanos into our XML-based monitoring system Specto and the lightweight, distributed event-based middleware Mitto. Our cooperative approach does not conflict with the decentralized parallelization algorithm of the distributed SAT prover. Empirical results show that our approach can contribute to the performance of distributed symbolic computation. In this example, a load balancing subsystem is implemented taking the now available global information about the system state appropriately into account.