Impact of Memory on Deep Learning Performance and LLM Feedback

·Dec 05, 2025 04:42 PM·

Priyan J. When it comes to prompt learning, does the performance increase when you add memory to the system? Like the "judge" LLM that assigns English feedback, does it get better if you add conversation history, or does it do better when it's truly like reinforcement learning?

2 comments

· Sorted by Oldest

Priyan J.
·
·
Hi Nouamane great question. If you are trying to optimize a multi-turn conversation, the judge LLM would definitely benefit from looking at conversation history. The best way to do this is to pass in traces into the meta prompt. That way, the meta prompt LLM (and judge LLM) have access to the entire path taken by the agent (tool calls, retrievals, all LLM calls) and can optimize the prompt with respect to that
🙌1
Nouamane
·
·
That makes sense, thanks. I will probably be making a PR to the repo soon FYI. I want it to be a cli and have the ability to optimize image generation too

Priyan J.
·
·
Hi Nouamane great question. If you are trying to optimize a multi-turn conversation, the judge LLM would definitely benefit from looking at conversation history. The best way to do this is to pass in traces into the meta prompt. That way, the meta prompt LLM (and judge LLM) have access to the entire path taken by the agent (tool calls, retrievals, all LLM calls) and can optimize the prompt with respect to that
🙌1
Nouamane
·
·
That makes sense, thanks. I will probably be making a PR to the repo soon FYI. I want it to be a cli and have the ability to optimize image generation too