The article deals with self-customizing software, and investigates kinds of patterns that can be recognized by the machine learning algorithm. A self-customizing software should be able to recognize stereotypical sequences of action, like steps involved in sending electronic-mail. Many researchers have developed methods that predict user's actions solely based on user's history of past interactions with a computer. Methods that look solely at the user's single immediately preceding action, comparing it to past situations where the user previously took that action, often give rise to evenhanded predictions. The author works on the Incremental Probabilistic Action Modeling (IPAM) prediction method, which learns to predict actions of a user interacting with the Unix command-line shell. IPAM predicts by simply recording commands observed, and maintains a probability distribution over all command that may follow. IPAM focuses on user's last interaction, paying more attention to the recent past, through its scaling down of probabilities of those actions not taken at the current point in time.