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?
Comments
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?