Can LLM agents improve without fine-tuning?

Yes. A memory-driven approach lets LLM agents store critiques from past tasks and retrieve them when similar problems come up, enabling continual learning without any parameter updates.