Adaptive behavior and learning are required of software agents in many application domains. At the same time agents are often supposed to be resource-bounded systems, which do not consume much CPU time, memory or disk space. In attempt to satisfy both requirements, we propose a novel framework, called APS (standing for Analysis of Past States), which provides agent with learning capabilities with respect to saving system resources. The new solution is based on incremental association rule mining and maintenance. The APS process runs periodically in a cycle, in which phases of agent's normal performance intertwine with learning phases. During the former ones an agent stores observations in a history. After a learning phase has been triggered, the history facts are analyzed to yield new association rules, which are added to the knowledge base by the maintenance algorithm. Then the old observations are removed from the history, so that in the next learning runs only recent facts are processed in search of new association rules. Keeping the history small can save both processing time and disk space as compared to batch learning approaches.