In the last decade, the availability of new minimally invasive subcutaneous sensors for monitoring glucose level continuously stimulated research on new online strategies for improving the treatment of diabetes, including hyper/hypoglycemic alert generators and artificial pancreas. An important aspect that has to be dealt with in these applications is the random measurement noise that affects continuous glucose monitoring (CGM) signals. One major difficulty is that for a given sensor technology, the signal-to-noise ratio (SNR) can vary from subject to subject (interindividual variability) and also within subject (intraindividual variability). Recently, a denoising approach implemented through a Kalman filter with parameters automatically tuned, once for all, in a burn-in interval was proposed to cope with the interindividual variability of SNR. In this paper, we propose a new denoising method able to cope also with the intraindividual variability of the SNR. The method resorts to a Bayesian smoothing procedure that uses a statistically-based criterion to determine, and continuously update, filter parameters in real time. The performance of the method is assessed on both Monte Carlo simulation and 24 real CGM time series obtained with the Glucoday system (Menarini, Florence, Italy). The method has a general applicability, also outside from the CGM context.