We are interested in technical issues in building multi-agent systems that use a distributed collective memory of previous problem-solving episodes. The basic idea is that a group of heterogeneous agents solve some collective problem. The resulting episode traces are divided up and stored in the collective memory of the problem-solving agents. In future episodes, these traces can be used, and adapted, so as to serve as a basis for improved performance in a similar episode of problem-solving. The notion of storing solutions to previously solved problems is consistent with other models of case-based learning, as well as models of learning based on chunking (e.g., Laird, Newell, & Rosenbloom, 1986).
In this paper, we will describe the techniques and principles we are currently using to prepare the data before storage into collective memory. The current status of our system is that we have several running examples of multiple agents solving problems (in a simulated domain called MOVERS-WORLD) by adapting previous traces of activity. We have not conducted tests that quantify the impact of data preparation on either the efficiency of stored traces or overall community performance. Appendix C contains more information on performance, efficiency, and evaluating the system.