Substantial evidence suggests that working memory (WM) leverages relational representations to provide flexible support for cognitive functions, a capacity likely derived from the dynamic nature of neural codes in WM. However, how these dynamic codes represent and maintain relations remains unclear. Here, we examined the transformation of neural geometries in the dorsal prefrontal cortex of monkeys performing a visuospatial delayed-match/nonmatch task, where the monkeys were instructed to hold the spatial location of a white square in WM to match it with the spatial location of a subsequent square. We found that the sensory manifold during the square's presence and the mnemonic manifold after the square's offset both aligned with the stimulus manifold. However, significant differences emerged between the sensory and mnemonic manifolds, exhibiting little correlation in their neural geometries. Further analysis on the dynamic transformation from the sensory to mnemonic manifold revealed a process of expansion followed by flattening: the asymmetric sensory manifold first expanded into a symmetric neural geometry immediately after the square's onset offset, which then gradually flattened along dimensions different from those initially expanded, culminating in an asymmetric mnemonic manifold. This dynamic process of reconstruction not only remained its faithfulness to the stimulus geometry but also gained the flexibility to meet task demands. In sum, this transformation from asymmetry to symmetry and back to asymmetry in neural geometry precisely illustrates the dynamics of memory reconstruction, shedding lights on the subjective nature of WM that generates both accurate and illusory representation of the world we lived in.