FORGE: Self-Evolving Agent Memory With No Weight Updates via Population Broadcast
The paper introduces FORGE, a novel protocol for enhancing decision-making in LLM agents through self-generated memory without requiring weight updates. It demonstrates significant improvements in performance across various models in a network-defense scenario. The findings suggest that FORGE can effectively reduce major failure rates and may help weaker models perform better.
- ▪FORGE stands for Failure-Optimized Reflective Graduation and Evolution and utilizes a population-based approach for memory evolution.
- ▪The protocol showed an improvement in average evaluation return by 1.7-7.7 times over zero-shot learning and by 29-72% over the Reflexion baseline.
- ▪Population broadcast was identified as a critical mechanism for performance gains, while graduation primarily conserves computational resources.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Artificial Intelligence arXiv:2605.16233 (cs) [Submitted on 15 May 2026] Title:FORGE: Self-Evolving Agent Memory With No Weight Updates via Population Broadcast Authors:Igor Bogdanov, Chung-Horng Lung, Thomas Kunz, Jie Gao, Adrian Taylor, Marzia Zaman View a PDF of the paper titled FORGE: Self-Evolving Agent Memory With No Weight Updates via Population Broadcast, by Igor Bogdanov and 5 other authors View PDF HTML (experimental) Abstract:Can LLM agents improve decision-making through self-generated memory without gradient updates? We propose FORGE (Failure-Optimized Reflective Graduation and Evolution), a staged, population-based protocol that evolves prompt-injected natural-language memory for hierarchical ReAct agents.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
