Toward an Improved Artificial Theory of Mind: Context, Memory, Dialogue, and Population-Scale Learning in Social Prediction

Large language models have renewed interest in artificial theory of mind, particularly the possibility that artificial intelligence (AI) systems can infer and respond to human beliefs, intentions, emotions, goals, and perspectives. Building on Matta (2026), a Joanna Briggs Institute scoping review that addressed whether AI should incorporate theory-of-mind capabilities and which architectures support implementation, this article moves from synthesis to architectural specification. It develops the CPMD-ToM framework (Context, Population-Scale Learning, Memory, and Dialogue), which proposes that improved artificial theory of mind depends on the integration of four sources of social information: population-scale pattern recognition, multimodal context, interaction memory, and present dialogue. Population-scale learning provides broad social priors; multimodal context provides immediate evidence from the present social field; interaction memory supplies diachronic continuity across encounters; and present dialogue provides the current intentional frame. These sources are organized through a neuro-symbolic layer that structures prediction around beliefs, goals, intentions, emotions, knowledge states, perspectives, and uncertainty. The framework applies a graded formalizability schema that distinguishes strongly formalizable, partly formalizable, and ethically protected components, and argues that artificial theory of mind requires ethical meta-awareness to prevent overclaiming, stereotyping, manipulation, privacy intrusion, and reduction of the user to a fixed profile. The paper contributes to psychology and AI by reframing the artificial theory of mind as a layered architecture for responsible social prediction rather than a static linguistic capacity, and by deriving a falsifiable hypothesis suitable for empirical ablation testing.