Some people have a directory that they fill up with markdown documentation for LLMs that they either write themselves or have an LLM write. Myself, I tend to write comments in files that have a pivotal role, for instance I have a special style of test that uses certain helper functions and the theory and practice of those tests is discussed there.
I have the experience that the context goes bad in an agent if you use it long enough. The prompts start out like:
Do A.
That’s really good but there is problem B
But if you use it long enough it seems to start going in circles and it would be a big waste of time if you think it will do the right thing if you argue with it enough except it never does. My answer is to start a fresh session and maybe cut and paste some text from the old session into the new session or into comments.
If any long term memory solution carries this poison state with it it won’t really be helpful.
Some people have a directory that they fill up with markdown documentation for LLMs that they either write themselves or have an LLM write. Myself, I tend to write comments in files that have a pivotal role, for instance I have a special style of test that uses certain helper functions and the theory and practice of those tests is discussed there.
I have the experience that the context goes bad in an agent if you use it long enough. The prompts start out like:
But if you use it long enough it seems to start going in circles and it would be a big waste of time if you think it will do the right thing if you argue with it enough except it never does. My answer is to start a fresh session and maybe cut and paste some text from the old session into the new session or into comments.If any long term memory solution carries this poison state with it it won’t really be helpful.
[dead]