We review the present state of our understanding of mantle convection with respect to geochemical and geophysical evidence and we suggest a model for mantle convection and its evolution over the Earth's history that can reconcile this evidence. Whole-mantle convection, even with material segregated within the D" region just above the core-mantle boundary, is incompatible with the budget of argon and helium and with the inventory of heat sources required by the thermal evolution of the Earth. We show that the deep-mantle composition in lithophilic incompatible elements is inconsistent with the storage of old plates of ordinary oceanic lithosphere, i.e. with the concept of a plate graveyard. Isotopic inventories indicate that the deep-mantle composition is not correctly accounted for by continental debris, primitive material or subducted slabs containing normal oceanic crust. Seismological observations have begun to hint at compositional heterogeneity in the bottom 1000 km or so of the mantle, but there is no compelling evidence in support of an interface between deep and shallow mantle at mid-depth. We suggest that in a system of thermochemical convection, lithospheric plates subduct to a depth that depends - in a complicated fashion - on their composition and thermal structure. The thermal structure of the sinking plates is primarily determined by the direction and rate of convergence, the age of the lithosphere at the trench, the sinking rate and the variation of these parameters over time (i.e. plate-tectonic history) and is not the same for all subduction systems. The sinking rate in the mantle is determined by a combination of thermal (negative) and compositional buoyancy and as regards the latter we consider in particular the effect of the loading of plates with basaltic plateaux produced by plume heads. Barren oceanic plates are relatively buoyant and may be recycled preferentially in the shallow mantle. Oceanic plateau-laden plates have a more pronounced negative buoyancy and can more easily founder to the very base of the mantle. Plateau segregation remains statistical and no sharp compositional interface is expected from the multiple fate of the plates. We show that the variable depth subduction of heavily laden plates can prevent full vertical mixing and preserve a vertical concentration gradient in the mantle. In addition, it can account for the preservation of scattered remnants of primitive material in the deep mantle and therefore for the Ar and (3)He observations in ocean-island basalts.