A continuous optimal filtering is suggested which is suitable for systems with unknown and nonstationary dynamic and measurement noises. On-line adjustment of the gain towards the optimal value is based on filtered autocorrelation functions of the innovation sequence.