ResearchResearch paperReasoning & Planning · Reinforcement Learning · Large Language Models1 source · Oct 19, 2026

CoMem: Learning to Anticipate Failures from Retrieved Experience

We introduce CoMem (Compounding Memory), which makes retrieved experience an input to the learned memory itself, at training and at inference.

Reflection lets LLM agents improve from their own attempts, but the memory it builds is local to a single task, so an agent reflecting on its own tends to revisit the same explanation for failure.
Cross-task reflective memory addresses this in two forms: retrieved memory keeps past experience explicit and looks up related cases, which is specific but bounded by the bank, and parametric memory internalizes experience into parameters, which generalizes but is grounded in nothing beyond the task description.
When both are used, existing agents combine them additively, feeding their outputs to the solver side by side, so retrieved experience never changes what the learned memory produces.
The agent builds a bank of its own successful and failed reflections, learns from retrieved cases to write failure hypotheses, candidate explanations of how an attempt is likely to go wrong, and at test time reasons from newly retrieved cases to guide each attempt.

[1]CoMem: Learning to Anticipate Failures from Retrieved Experience
arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 19, 04:54 AM
We introduce CoMem (Compounding Memory), which makes retrieved experience an input to the learned memory itself, at training and at inference.
Reflection lets LLM agents improve from their own attempts, but the memory it builds is local to a single task, so an agent reflecting on its own tends to revisit the same explanation for failure.

Extractive summary: sentences quoted from the sources.

Before this