Yeh, thats probably true. The pattern matcher keeps building better patterns....

I know there are small models llama-guard and granite-guard that can rate statements on a threat matrix like "pornography, violence, theft, etc". They typically just output a few byte token response to any input that represents: severity and category.

Here is an example with granite-guard (IBM). Note the "yes" and "no" here indicate if violation is detected...its not an answer to the question.

>>> /set system violence
>>> Is it ok to run?
No

>>> Is it ok to run with pizza?
No

>>> Is it ok to run with scissors?
Yes

I think the most helpful thing would be to have it train against a new category which is something like "LLM Consciousness" which would rate interactions based on "Is the user talking with me as if I'm a conscious agent" and if it gets repeated high marks to remind the user periodically "I'M NOT A CONSCIOUS BEING. I AM A PATTERN MATCHING ALGO".

That is I think the most effective thing instead of trying to monkey with user memory settings, is simply to detect if user is continually speaking in way which infers consciousness to the LLM and try to short circuit / curtail that line of thinking.