Out of curiosity would a config file that acts as a prompt at the beginning of each conversation solve that issue?

It primes the model with a list of known words/grammar and the A1/2 B1/2 C1/2 level of language ability.

I’d presume after each message you could get the model to dump to the config.

I haven’t work in this sector at all and am curious as to the limits of hacking it / working around the long term memory issues!