Arthur’s paradigm of the El Farol bar for modeling bounded rationality and inductive behavior is undertaken. The memory horizon available to the agents and the selection criteria they utilize for the prediction algorithm are the two essential variables identified to represent the heterogeneity of agent strategies. The latter is enriched by including various rewarding schemes during decision making. Though the external input of comfort level is not explicitly coded in the algorithm pool, it contributes to each agent’s decision process. Playing with the essential variables, one can maneuver the overall outcome between the comfort level and the endogenously identified limiting state. The distribution of algorithm clusters significantly varies for shorter agent memories. This in turn affects the long-term aggregated dynamics of attendances. We observe that a transition occurs in the attendance distribution at the critical memory horizon where the correlations of the attendance deviations take longer time to decay to zero. A larger part of the crowd becomes more comfortable while the rest of the bar-goers still feel the congestion for long memories. Agents’ confidence on their algorithms and the delayed feedback of attendance data increase the overall collectivity of the system behavior.