In traditional way, software plans are represented explicitly by some semantic schemas. However, semantic contents, constrains and relations of plans are hard for explicit presentation. Besides, it is a heavy and error-prone work to build such a library of plans. Algorithms of recognition of such plans demand exact matching by which semantic denotation is obvious itself. We thus present a novel approach of applying neural network in the presentation and recognition of plans via asymmetric Hebbian plasticity and Non-linear Auto-Regressive with eXogenous inputs (NARX) to learn and recognize plans. Semantics of plans are represented implicitly and error-tolerant here. The recognition procedure is also error-tolerant because it tends to match fuzzily like human. Models and relevant limitations are illustrated and analyzed in this article.