Teaching agents to learn from failure without retraining

FORGE addresses the challenge of helping LLM agents improve performance through experience without gradient-based fine-tuning. The approach uses a two-level system: an inner reflection loop converts failed trajectories into reusable knowledge (textual rules, few-shot examples, or both), while an outer loop shares the best-performing memory across a population of agents between training stages. Evaluated on CybORG CAGE-2, a network-defense task, FORGE shows 1.7–7.7× improvement over zero-shot and 29–72% over standard Reflexion across four LLM families. Population broadcast—not memory freezing—drives the gains. Examples-based memory works best overall; rule-based memory offers better efficiency. Code and results are anchored to a single benchmark.