I don't know the solution to your problem Giorgio, but will offer a work-around that may help.
When I need to do a composite scene I do the following:
Create the scene empty of the figures, render it
Be sure at this point to lock all layers so nothing gets accidently moved
If there are areas of the scene where no figure will be loaded, and they do not effect the shadows or lighting, then you could delete those areas/items to save resources
Add a figure and then select a render box closely around the figure - being sure to include any shadows or reflections in nearby objects, render the selected smaller area
Delete the first figure and load the second, then repeat the process of selecting and rendering the effected area.
The result is one large full scene image and a series of smaller images, since this has all been done without closing the original scene, as long as the camera doesn't get moved your POV should stay constant.
You can then merge them all together in Photoshop or another graphics program.
I also use this when I find a small problem after a render, such as poke through or floating objects. I fix the problem and then just render the local area round it, saving lots of time compared to re-rendering the whole image.