FORGE: Self-Evolving Agent Memory via Population Broadcast — No Weight Updates Required

arXiv·medium signal

FORGE (Failure-Optimized Reflective Graduation and Evolution) introduces a staged, population-based protocol that evolves prompt-injected natural-language memory for hierarchical ReAct agents. A dedicated reflection agent converts failed trajectories into reusable knowledge artifacts, which are broadcast across a population of agents. Unlike fine-tuning approaches, FORGE uses the same underlying LLM with no distillation from a stronger model — memory improvements come purely from structured reflection on failures.