We examine an application of the optimal prediction framework to the truncated Fourier-Galerkin approximation of Burgers's equation. Under particular conditions on the density of the modes and the length of the memory kernel, optimal prediction introduces an additional term to the Fourier-Galerkin approximation which represents the influence of an arbitrary number of small wavelength unresolved modes on the long wavelength resolved modes. The modified system, called the t-model by previous authors, takes the form of a time-dependent cubic term added to the original quadratic system. Numerical experiments show that this additional term restores qualitative features of the solution in the case where the number of modes is insufficient to resolve the resulting shocks (i.e., zero or very small viscosity) and for which the original Fourier-Galerkin approximation is very poor. In particular, numerical examples are shown in which the kinetic energy decays in the same manner as in the exact solution, i.e., as t −2 when t shock t Re, even when a very small number of resolved modes is used. Correlation-like quantities related to the memory kernel are then computed, and these exhibit a t −3 tail for the same time period.1. Introduction. The optimal prediction framework developed by Chorin and coworkers (see, for instance, [8,10,19,9,6]) can, among other things, be regarded as a type of order reduction scheme. Such schemes attempt to take a very large dynamical system and reduce its size by taking into account the effects of a large subset of the variables (which we will call unresolved) on a small subset (resolved) without explicitly computing their evolution. This large subset is usually thought of as consisting of relatively uninteresting variables at the extreme end of a scale, e.g., fast or small. Loosely speaking, optimal prediction includes these effects by the addition of an integral term, called the memory term, to the original set of equations for the resolved variables. However, the kernel in the memory term is difficult to compute, paradoxically much more difficult than a direct solution of the full system. Hence in practice it is customary to approximate it, and the typical approximation is that its support is short; i.e., the evolution of the resolved variables depends on their current state at time t and the influence of the unresolved variables at t and perhaps for a very short time in the past. For certain systems this short memory approximation is appropriate [3]. There is, however, evidence that there are some systems, in particular fluids, for which the memory effect is long [1,2], and based on this Chorin and Stinis [9] have conjectured that the Navier-Stokes equations have a long-time memory kernel. Therefore, Burgers's equation is a good place to start when looking for such long memory effects in continuum models of fluid systems.Another motivation for this work is the numerical method know as large eddy