Working memory refers to information actively held in the brain and is essential for advanced functions like thinking, decision-making, or learning. In this thesis, I focus on two key problems in working memory: 1) how to maintain a graded amplitude of a local memory in a classical ring architecture, and 2) how to maintain a novel pattern of graded neural activity in an unstructured network. To address these problems, I start with reexamining previous methodologies to limit modeling possibilities. Then, I propose two neural circuit models with dendritic bistability, each of which is treated analytically and which are robust against various perturbations. This work links physiological properties to functionality, contributing to a deeper understanding of the mechanisms underlying working memory.