How we feel reflects a combination of recalled and recognized emotions. All existing self-report measures are based solely on recognized emotions. To understand the influence of recalled emotions, we developed a new method to recover human emotional states based on emotional free association, in a task we call the emotional fluency. The present work investigated the differences between recall and recognition in human emotional states. We compared the emotional fluency task with self-report measures, including PANAS, WEMWBS, and the Emotional Intelligence Scale. Using language statistics computed from the emotional fluency task, we developed multiple models for predicting self-report measures. We find that while recalled emotions can predict recognized emotions, they highlight important problems with existing recognition measures, including emotional coverage and the difference between availability and accessibility. We also investigate the search process in emotional memory, supporting the role of unbiased memory sampling and higher emotional intelligence and mental well-being.