The article proposes a comprehensive framework, the theory of probabilistic mental models (PMM theory), to explain people's confidence in their general knowledge. PMM theory explains two stable effects: the overconfidence effect (mean confidence is higher than the percentage of correct answers) and the hard-easy effect (overconfidence increases with item difficulty). It also predicts a new phenomenon, the confidence-frequency effect, which is the systematic difference between confidence in a single event and frequency judgments of correct answers over time. The theory suggests that people construct local mental models (LMs) for simple tasks and probabilistic mental models (PMMs) for more complex tasks. LMs are based on memory and elementary logical operations, while PMMs use probabilistic information from a natural environment. The theory predicts that in typical general-knowledge tasks, people show overconfidence and accurate frequency judgments, but in randomly sampled tasks, overconfidence disappears and frequency judgments underestimate true frequencies. Two experiments support these predictions and explain apparent anomalies in previous research. The theory provides a unified framework for understanding confidence and frequency judgments, integrating various experimental findings and explaining their inconsistencies.The article proposes a comprehensive framework, the theory of probabilistic mental models (PMM theory), to explain people's confidence in their general knowledge. PMM theory explains two stable effects: the overconfidence effect (mean confidence is higher than the percentage of correct answers) and the hard-easy effect (overconfidence increases with item difficulty). It also predicts a new phenomenon, the confidence-frequency effect, which is the systematic difference between confidence in a single event and frequency judgments of correct answers over time. The theory suggests that people construct local mental models (LMs) for simple tasks and probabilistic mental models (PMMs) for more complex tasks. LMs are based on memory and elementary logical operations, while PMMs use probabilistic information from a natural environment. The theory predicts that in typical general-knowledge tasks, people show overconfidence and accurate frequency judgments, but in randomly sampled tasks, overconfidence disappears and frequency judgments underestimate true frequencies. Two experiments support these predictions and explain apparent anomalies in previous research. The theory provides a unified framework for understanding confidence and frequency judgments, integrating various experimental findings and explaining their inconsistencies.