The national-scale 'repères de crues' database ('flood marks', www.reperesdecrues.developpement-durable.gouv.fr) contains thousands of geolocalized historical flood marks spanning several centuries. This constitutes a valuable source of information complementing data from the network of hydrometric stations in order to characterize ancient flood events and, possibly, to improve flood hazard assessment. However, this information is not obvious to use, in particular because the network of flood mark sites is distinct from the network of hydrometric stations. The aim of this work is to address this issue by means of a probabilistic model jointly describing the occurrence of flood marks at sites and annual maximum flood peaks at stations. A Bernoulli distribution is used for the former and a Generalized Extreme Value distribution for the latter. Some parameters of these distributions are allowed to vary in time following a set of unobserved latent time series called Hidden Climate Indices (HCI). Importantly, the same HCIs drive the temporal evolution of both flood marks and peaks: this allows transferring information between the two variables despite them being measured on distinct networks. The model is applied to about 300 flood mark sites (1707-2015) and 200 hydrometric stations (1904-2015) in France. Results illustrate how the information carried by flood marks at sites can be used to estimate the time-varying probability of exceeding some high discharge threshold at stations during the whole period 1707-2015, which largely predates the existence of stations. In general, these estimations are sharp in the Oceanic part of the country (probabilities may approach 0 or 1), much less in the Mediterranean area (probabilities remain near-constant). Probability maps can also be derived and provide a quantitative information on the extent and spatial structure of ancient flood events. The model also allows deriving the marginal distribution used in Flood Frequency Analysis, but results indicate that the model does not make an efficient use of flood marks to improve the estimation of this distribution. Overall, this work illustrates the interest of deriving flexible probabilistic models to make the best use of several existing sources of information and data while adapting to their varied characteristics.