The article explores the computational foundations of dynamic coding in working memory (WM), a fundamental aspect of cognition. Traditional views suggest that WM maintenance relies on stable neural activity patterns, but recent evidence indicates that neural activities during WM maintenance exhibit dynamic variations before stabilizing. The authors review the evidence for dynamic coding in prefrontal cortex (PFC) recordings, showing that neural activities change significantly during the cue and early delay periods but stabilize during the late delay period. They also examine the dynamics of classical and task-optimized neural network models, finding that task-optimized models naturally exhibit dynamic coding, while classical models do not. The authors propose that dynamic coding results from an optimality principle, enhancing task performance across various constraints and network architectures. They suggest that dynamic coding is a fundamental computational feature of WM maintenance, rather than an epiphenomenon. The article concludes by discussing the generality of dynamic coding and the factors that influence its strength, such as task complexity and network connectivity.The article explores the computational foundations of dynamic coding in working memory (WM), a fundamental aspect of cognition. Traditional views suggest that WM maintenance relies on stable neural activity patterns, but recent evidence indicates that neural activities during WM maintenance exhibit dynamic variations before stabilizing. The authors review the evidence for dynamic coding in prefrontal cortex (PFC) recordings, showing that neural activities change significantly during the cue and early delay periods but stabilize during the late delay period. They also examine the dynamics of classical and task-optimized neural network models, finding that task-optimized models naturally exhibit dynamic coding, while classical models do not. The authors propose that dynamic coding results from an optimality principle, enhancing task performance across various constraints and network architectures. They suggest that dynamic coding is a fundamental computational feature of WM maintenance, rather than an epiphenomenon. The article concludes by discussing the generality of dynamic coding and the factors that influence its strength, such as task complexity and network connectivity.