Learning enables animals to construct flexible knowledge structures, allowing them to adapt to new tasks by reconfiguring prior experiences into novel strategies. This compositional nature of learning suggests that fundamental building blocks from past experiences are essential for future problem-solving. To investigate how experience shapes learning, we trained Recurrent Neural Networks (RNNs) to model a complex, context-dependent odor timing task, using constraints derived from prior reports detailing mouse behavior and shaping procedures. The task, a temporal variant of the Delay Non-match to Sample (tDNMS), involves timing the duration of two stimuli and then identifying trials of non-matching durations. Non-match (Long-Short, Short-Long) and match (Short-Short) trials constitute distinct temporal contexts and result distinct behaviors (Go vs No-Go) during a response window. Learning this task depends on the medial entorhinal cortex (MEC), a neural subregion where local recurrent circuits are thought to comprise an attractor network generating periodic activity during navigation and, potentially, during timing. During timing two key questions remain unanswered: what mechanisms permit network activity to traverse context specific trajectories? and how is this process influenced by past experience? Our results demonstrate that pre-training with shaping tasks improves RNNs’ eventual performance; further we reveal a mechanism for this effect. Networks without shaping commit incorporate stereotypical behavioral errors, indicating that they fail to properly distinguish the temporal contexts. Shaping cultivates subtle but significant alterations in network connectivity, which result in characteristic differences observed in network population activity. In this way, shaping tasks allow RNNs to develop an abstract representation of timing. The selection of shaping task is important: training on a single trial type (Long-Short or Short-Long) instead of both impairs future learning. Finally, electrophysiological recordings reveal identical time-coding axes across trial types, indicating that mice with high performance learn the same abstraction as the RNNs.