Many AI problems can be modeled as constraint satisfaction problems (CSP), but many of them are actually dynamic: the set of constraints to consider evolves because of the environment, the user or other agents in the framework of a distributed system. In this context, computing a new solution from scratch after each problem change is possible, but has two important drawbacks: inefficiency and instability of the successive solutions. In this paper, we propose a method for reusing any previous solution and producing a new one by local changes on the previous one. First we give the key idea and the corresponding algorithm. Then we establish its properties: termination, correctness and completeness. We show how it can be used to produce a solution, either from an empty assignment, or from any previous assignment and how it can be improved using filtering or learning methods, such as forward-checking or nogoodrecording. Experimental results related to efficiency and stability are given, wi...