In this study, a dynamic modeling and real-time intervention system based on deep reinforcement learning is proposed to address the problems that static models in current children’s cognitive development intervention are difficult to dynamically track changes in individual cognitive states and that personalized strategies are not adaptable enough. The cognitive state space is constructed by collecting multimodal behavioral data (including eye tracking, task completion time, and EEG signals) of children aged 3–6 years old, and an improved deep Q-network (DQN) algorithm is used to establish a dynamic model of cognitive development, which defined a three-dimensional state vector including attention level (0–100), working memory capacity (1–5 units), and executive function score (0–10). A dual-channel reward function (immediate task reward + long-term cognitive development reward) is designed, and transfer learning is introduced to use the 1200 historical cognitive development data pre-trained model as the initial parameters. The experimental results show that in a 6-month longitudinal experiment involving 150 children, after the experimental group (n = 75) adopted the system intervention, the standardized cognitive assessment scale score increases to 92.3, and the attention task completion time is reduced to 32.7 s, verifying the effectiveness of the system in dynamically modeling cognitive trajectories and generating personalized intervention strategies, providing a scalable intelligent solution for children’s cognitive development monitoring.