Abstract. We consider time-dependent parabolic problems coupled across a common interface which we formulate using a Lagrange multiplier construction and solve by applying a monolithic solution technique. We derive an adjoint-based a posteriori error representation for a quantity of interest given by a linear functional of the solution. We establish the accuracy of our error representation formula through numerical experimentation and investigate the effect of error in the adjoint solution. Crucially, the error representation affords a distinction between temporal and spatial errors and can be used as a basis for a blockwise time-space refinement strategy. Numerical tests illustrate the efficacy of the refinement strategy by capturing the distinctive behavior of a localized traveling wave solution. The saddle point systems considered here are equivalent to those arising in the mortar finite element technique for parabolic problems. 1. Introduction. We consider an adjoint-based a posteriori error estimator for parabolic problems that are coupled across a given interface and construct an adaptive space-time finite element scheme. A posteriori error estimation is commonly used in parabolic problems for adaptive mesh refinement and the different approaches are described in a detailed review paper [1], as well as in [4,5,14,20,24,35,39,40,41]. In particular, [14,15,16] develop a technique based on solving an adjoint problem to estimate the error in a quantity of interest given by a functional of the solution as opposed to the error in a norm of the solution.Error estimation for coupled problems solved using operator decomposition techniques was considered in [17,18,9,8]. While operator decomposition is traditionally viewed as easy and inexpensive to implement [19,28], its convergence is rarely guaranteed. In this paper we take the alternative, monolithic approach to solving coupled parabolic problems, which do not suffer from this problem. We introduce a Lagrange multiplier (mortar element) space on the interface between the component domains, resulting in a large-scale complex system, which can nevertheless be solved efficiently with special treatment [33,22]. Moreover, the presence of the Lagrange multiplier variable allows us to derive error estimates for a quantity of interest that is supported along the interface. This is particularly useful in applications where the Lagrange