Working memory (WM) refers to the part of short-term memory that temporarily holds and processes information when it is no longer accessible to the senses. It relies on briefly storing stimulus features in the activity of neuronal populations. Despite its crucial role in cognitive functions, WM is especially vulnerable due to its limited capacity and stability, especially when facing distractions from internal and external sources. Recent studies have explored the neural mechanisms that safeguard WM. It is proposed that pre- and post-distraction neural activity decomposes into orthogonal subspaces, thus protecting information. However, whether orthogonalization is innate or acquired through learning is unknown, and the network mechanisms supporting it are unclear. Here, we probe WM orthogonalization using calcium imaging data from the mouse prelimbic (PrL) and anterior cingulate (ACC) cortices as they learn to perform an olfactory dual task. The dual task combines an outer Delayed Paired-Association task (DPA) with an inner Go-NoGo task. We examined how neuronal activity reflected the process of shielding the DPA sample information against Go/NoGo distractors. As mice learned the task, we measured the overlap between the neural activity onto the low-dimensional subspaces that encode sample or distractor odors. Early in the training, pre-distraction activity overlapped with both sample and distractor subspaces. Later in the training, pre-distraction activity was strictly confined to the sample subspace, resulting in a more robust sample code. We present a mechanistic insight into how these low-dimensional WM representations evolve with learning in a recurrent neural network model of excitatory and inhibitory neurons with low-rank connections. The model links learning to (1) the orthogonalization of sample and distractor WM subspaces and (2) the orthogonalization of each subspace with irrelevant inputs. We validated (1) by measuring the angular distance between the sample and distractor subspaces through learning in the data. Prediction (2) was validated in PrL through the photoinhibition of ACC to PrL inputs, which turned back late training into early training dynamics. Moreover, our model constrains the network dynamics to a double-well attractor on a one-dimensional ring and suggests that learning optimally adjusts the location of the attractors on this ring. We validated this theoretical prediction by estimating an energy landscape for the recorded neural dynamics. In sum, our study underscores the crucial role attractor dynamics plays in shielding WM representations from distracting tasks.