FORGE: LLM Agent Memory Evolution via Population Broadcast

A new protocol called FORGE (Failure-Optimized Reflective Graduation and Evolution) enables LLM agents to improve decision-making through self-generated memory without weight updates. It uses a population-based approach where prompt-injected natural-language memory evolves across stages. An inner Reflexion loop converts failed trajectories into reusable artifacts (rules, examples, or mixed), while an outer loop propagates the best memory across the population and freezes converged instances. Tested on CybORG CAGE-2, a stochastic network-defense POMDP, with four LLM families (Gemini-2.5-Flash-Lite, Grok-4-Fast, Llama-4-Maverick, and others), FORGE demonstrates performance gains without gradient updates.