A new paper on arXiv introduces CoMem, a method for helping large language model agents anticipate failures by drawing on retrieved experience. The authors point to a known limitation of reflection: when an agent improves from its own attempts, the memory it builds is local to a single task. As a result, the agent often ends up revisiting the same explanation for why it failed, rather than generalising from that experience.
CoMem is positioned as a response to that problem, shifting from task-local reflection to cross-task reflective memory. The idea is that an agent can anticipate failures more effectively when it retrieves and uses relevant experience from other tasks, instead of relying only on its own repeated attempts.
The abstract is brief and does not include experimental details or comparisons with other methods, so the reported benefits are not yet independently verified. Still, the core motivation is clear: reflection alone may be too narrow, and memory that spans tasks could be a useful direction for more robust LLM agents.