Personalized Vocabulary Memory Curve Optimization Model Based on Deep Reinforcement Learning

Stable vocabulary retention is essential in university English learning, but individual cognitive differences often produce unstable memory curves. This paper proposes a personalized memory optimization model based on deep reinforcement learning to construct adaptive review strategies. An LSTM network encodes time-series behavioral data, including accuracy, reaction time, and review time, into a 128-dimensional latent state vector that captures dynamic memory characteristics. Memory strength is mapped into a four-level discrete state space, including newly learned, weakened, consolidation, and stable states, while five optional review timings form the action space. Proximal Policy Optimization learns the optimal review policy by optimizing a reward function that balances retention gain against learning-load penalty. The closed-loop process continuously incorporates learning feedback. Experiments show the highest accuracy of 89.2%–93.1% with low review frequency of 4.3–5.8 times. Long-term retention is also improved, with the average vocabulary retention rate on day 60 not lower than 78.3% and the average memory fluctuation coefficient not higher than 0.18. The proposed method alleviates retention instability caused by cognitive differences and variable learning pace, providing a personalized sequential-decision framework for adaptive learning systems.