LLM agents can get better at tasks without updating their model parameters by saving their own past executions in external memory and reusing that experience to guide future decisions. This approach raises a central design question: what should the agent retain? The answer, according to a new paper on arXiv, is not simply "everything."
The paper, titled "More Memory, Worse Agents," reports that keeping too much memory can actually hurt performance. The authors describe two failure mechanisms: error reproduction, where mistakes from earlier runs are carried forward and repeated, and anti-persistence, where the agent overcorrects and switches strategies too readily, undermining consistent progress.
These findings suggest that memory management in LLM agents is a delicate trade-off. Rather than accumulating all experience, agents may need selective retention—keeping only what is useful and avoiding the pitfalls of excessive recall. The study does not yet offer a precise rule for what to keep, but it highlights a clear risk for anyone designing memory-augmented agents.