@krika2810 I'm not sure what the popular harnesses offer, but it's easy in a custom harness, and open source harnesses like hermes agent. You just need write access to the context sent to the model, and read access of the traces of its loops. I use a cheap qwen 30b model to handle context
@TinfoilTricorn awesome, I'm liking the idea of the looping transformer to save RAM without cutting flops. bonus points if relevant data can be injected into the loop too. Should help get more out of modest hardware.