We present a new method of incremental, statistical learning, which is suitable for knowledge-based systems, especially software agents. The method is based on the imperfect recall assumption, according to which an agent does not store all the past observations. However it does preserve general rules concerning the past, that can be potentially useful for improving agent?s action. During its performance an agent stores observations in the history. When system resources are idle and the size of the history is sufficient as for its statistical significance, the stored facts are analysed by means of data mining techniques, and disposed afterwards. The discovered rules are combined with the former rule base, so that the final rule set is approximately the same, as if it was obtained on the whole history.