so i think I've written a harness which accomplishes this
https://arxiv.org/abs/2609.37725
but using any model. one key aspect is that the model both generates and manages its own context but does so at the harness level without fine tuning.
its a unique approach and doesnt require external memory files but could include one but the hardwork is done as the model decodee and probably requires 2 threads to run without locking
if anyones interested I plan on benchmarking it tonight and cleaning it up to share #llm



