I use Burnless every day, and I really think that is really useful. And the most impressive thing is that the LLM learns how to use it instantly. It's really interesting to see how AI thinks when they are delegating or deciding about the /clear to help us.
Yesterday in a full real workday I used 1.44M tokens. But in context only
1,590 tokens. That's 908x less context and context can survive any /clear.
I'm not claiming that Burnless took the exact same 1.4M tokens of text and
compressed it to 1,590 tokens. What really happened was that all of that
work was done, recorded, verified and continued without the accumulated
history remaining inside the active context.
To repro with your own key: python bench/run.py --turns 10
I posted here to get technical criticism from people who use long LLM or
agent sessions.
Author here.
I use Burnless every day, and I really think that is really useful. And the most impressive thing is that the LLM learns how to use it instantly. It's really interesting to see how AI thinks when they are delegating or deciding about the /clear to help us.
Yesterday in a full real workday I used 1.44M tokens. But in context only 1,590 tokens. That's 908x less context and context can survive any /clear.
I'm not claiming that Burnless took the exact same 1.4M tokens of text and compressed it to 1,590 tokens. What really happened was that all of that work was done, recorded, verified and continued without the accumulated history remaining inside the active context.
To repro with your own key: python bench/run.py --turns 10
I posted here to get technical criticism from people who use long LLM or agent sessions.