Not exactly agents ignoring instructions, but I'm wondering if (selective) inclusion of LLM responses in a RAG stack might be suitable as a sort of long-term memory for "accepted" LLM contributions to code for example. This way, unwanted novel or alternate solutions to repeated patterns might be avoided?