This paper studies a direct approach to smoothing by sampling the posterior distribution in four dimensional data assimilation. The methodology is based on a hybrid Monte Carlo approach and can be applied to non-linear models, non-linear observation operators, and non-Gaussian probability distributions. The generated ensemble is used to construct both the analysis state (the minimum variance estimator) and the analysis error covariance matrix. Numerical tests performed with the Lorenz-96 model and with both linear and quadratic observation operators illustrate the usefulness and performance of the approach.