I was also suffering from a severe memory decline (halving over a period of a week after a reboot) but @lewis.heidrick's post here RM and a Reboot got me thinking about runcounts.
Turns out it was a maker API issue where I was sending MQTT data to node-red and then to a virtual device. A lot of the MQTT data was redundant but it kept refreshing the virtual device. Simply limiting the refresh to an appropriate rate and doing the same for all my MQTT links seems to have made a huge reduction in the loss of free memory.
So it really does highlight the need to optimise the IO to the hub wherever possible.
