The study of past climates and of mechanisms that have influenced their evolution is the key to anticipate the future climate changes. This doctoral thesis focusses on the Holocene climate, the ongoing interglacial, that starts about 11,700 years ago. The current paleoclimate knowledge is based on the one hand, on the climate models results and, on the other hand, on the reconstruction of physical variables derived from climate archives as the ice cores, the marine cores or the pollens for instance. These two types of information are complementary. Here we have combined them to obtain reconstructions of past climates using data assimilation. This technique is standard in many disciplines but not yet in paleoclimatology. The data assimilation method applied here is a particle filter. It is based on the selection of the members of an ensemble of simulations performed with the climate model LOVECLIM that have the best agreement with the reconstructions. Additional simulations are designed to investigate the contribution of various forcing to the observed changes over the Holocene. Our results demonstrate that the data assimilation can be used to objectively identify the incompatibilities that could exist between temperature reconstructions that are based on different archives. Also, the data assimilation methodology allows selecting mechanisms that, according to LOVECLIM, have influenced the climate as reconstructed using indirect indicators. For instance, about 2700 years ago, the colder Arctic climate suggested by the proxy data, may be the result of a decrease in westerlies. Including the climatic response to volcanic eruptions in our experiments does not significantly improve the agreement between simulated results and reconstructions over the Holocene.