A two-layered architecture for reasoning that uses narratives to guide its behavior is presented. The narratives are interpreted by reactive and deliberative reasoning layers to generate responses to external events, in accordance with internal desires created by the deliberative layer. A C++ implementation of the proposed Pickman architecture is described and tested in five scenarios in which three different Pickman versions are tasked with healing a small population suffering from a virtual plague. It is shown that a Pickman implementation using a static desire generation mechanism is efficient, but unreliable. The dynamic version, on the other hand, is not as efficient, but it performs successfully in all scenarios.