In the brain, massive interactions between neurons through synapses give rise to rich dynamics and have been thought to be critical for brain computation. In this talk, I will discuss recurrent network models for working memory that refers to an ability to maintain information on a time scale of seconds. Persistent neural activity in the absence of stimulus has been identified as a neural correlate of working memory, and it has been suggested that network interactions must be used to prolong the duration of persistent activity. Using dynamical systems theory and control theories, I found a new mechanism for generating persistent activity based on the principle of corrective feedback both in spatially homogeneous networks (Lim and Goldman, Nat. Neurosci., 2013) and in spatially structured networks (Lim and Goldman, J. Neurosci., in press). Several advantages of this new network model compared to previous models will also be discussed.