Hi Doc!
A strange behavior!
I've noticed that a long time ago but now I've decided to tell you.
It may help you debugging the code.
I've automated startup procedure for the miner whenever windows starts.
Let's take for example one of my rigs with 6 cards.
Windows boots and after a while miner starts.
After some time of stabilization watching memory usage in task manager I see it stable at ~286 - 290 M.
If I kill the miner process and start it again giving it time to stabilize it stays at ~560 - 570 M.
That's why I was saying all the time about "memory management".
I don't know but this is the behavior.
It may help.
Greetings!
ps: in heavy algo.
well if you would follow this thread regulary you would know that this is very well known, and can be fixed by just fireing up GPU-Z

You know something Doc, you didn't understand me.
I'm not talking about the known hashrate drop problem which I first introduced to the community the temporary solution with GPU-Z
but about different seized memory (MB) into 2 different runs of miner which of course may related.
No hard feelings!
No hard feelings, we did not understand eachother.
Also i don't take notes who said/wrote what and when , so i couldn't know you found this nice 'fix'.

Is this behaviour you are referring to happening every time (like first run always less mem used, second run always ok) ?
I see other miners have this 'problem' too on heavy.
Yes, happening all the time.
Miner uses about half the mem every time machine reboots and for every next run about double of that.
Also what means compute errors?
I've installed a new Gigabyte rx580 8gb card.
I have no memory errors, no hang ups, working fine but only this new card reports 1 or 2 compute errors in stats in a day!