A distributed adaptive filter can estimate an unknown signal of interest by a set of sensors working cooperatively, when any individual sensor cannot fulfill the filtering task due to lack of necessary information condition. This paper considers these kinds of filtering problems and focuses on a class of consensus normalized least mean squares-based algorithms. A general and weakest possible cooperative information condition is introduced to guarantee the stability of these kinds of adaptive filters, without resorting to commonly used but stringent conditions such as independence and stationarity of the system signals, which makes our theory applicable to feedback systems. Moreover, this general information condition is shown to be not only sufficient but also necessary for stability of the adaptive filters for a large class of random signals with decaying dependence. We further show that the mean square tracking error matrix can be approximately calculated by a linear deterministic difference matrix equation that can be easily evaluated and analyzed.