We describe a new method for making palaeoclimate reconstructions. The method produces seasonal palaeoclimate variables from site-based reconstructions, which are used to further generate reconstructions of six key variables: mean annual precipitation, mean annual temperature, mean temperature of the coldest month, mean temperature of the warmest month, growing degree days above 5â—¦C (a measure of the growing season) and moisture index (the ratio between precipitation and equilibrium evapotranspiration). The method uses a variational technique, 3D-Var, to produce the maximum a posteriori estimate of the climate given pollen-based reconstructions and a prior estimate. We apply spatial and temporal correlations in the error of the prior to ensure that the reconstruction is spatially and temporally smooth. The strength of the correlations can be adjusted with scaling parameters to determine the smoothness of the reconstruction. Further, the method uses an observation function that accounts for the difference in atmospheric CO2 concentration between the modern and palaeo time periods; this means the final reconstructions have the correct response to CO2 change. We explore the effect of the length scales on the reconstruction to determine the optimal scales. The method is applied to pollen-based reconstructions together with a prior estimate made up of outputs from the 3rd round of the Palaeoclimate Modelling Intercomparison Project and is used to make global reconstructions of the Last Glacial Maximum (c.a. 21,000 years ago) and the mid-Holocene (c.a. 6,000 years), as well as estimates of error for the reconstructions. The reconstructions show good spatial and temporal smoothness and produce realistic climates that can be used for data-model comparison.