How does the brain gain insight from limited novel experience? Humans and animals have a striking ability to learn relationships between items of experience, enabling structured generalization and rapid assimilation of new information. A fundamental type of such relational learning is order learning, which enables transitive inference: if A>B and B>C, then A>C. Remarkably, both humans and animals can quickly and globally reorganize separate learned orderings upon learning a single new item ("list-linking"; learn A>B>C and D>E>F, then C>D, and infer B>E), an instance of fast "reassembly" of existing knowledge. Yet despite longstanding study, neural mechanisms of transitive inference and fast knowledge reassembly have remained elusive. To address this problem, we adopted a meta-learning ("learning-to-learn") approach. We meta-trained neural networks, endowed with synaptic plasticity and neuromodulation, and thus capable of self-directed learning, to learn orderings of (arbitrary) stimuli from presentation of stimulus pairs, similarly to classic transitive inference experiments. We then obtained a complete mechanistic understanding of these discovered neural learning algorithms. We found that that two solutions consistently emerged. One learns orderings passively, via Hebbian association of current-trial stimuli with response, modulated by reward. This “passive” solution can perform transitive inference, but not list-linking. By contrast, the other solution uses an active mechanism: in each trial, relevant items from previous trials are selectively and actively reinstated in working memory, enabling delayed Hebbian learning to modify the representation of these past items (in addition to current-trial items). Crucially, this "active" solution is capable of list-linking (fast knowledge reassembly) and reproduces many observed behavioral features of transitive inference - to our knowledge, the first neural model to do so. These results identify a neural mechanism for relational learning, fully elucidate its mechanistic operation, and also highlight a novel approach (meta-learning with plastic networks) for discovering such mechanisms.