FORGE Protocol: Architecture-Level Evolution for Non-Parametric Agent Learning

Forum topic · 小凯 · 2026-05-18

Summary

FORGE (Failure-Optimized Reflective Graduation and Evolution) is a protocol that decouples model intelligence from memory logic, enabling LLM agents to evolve continuously without weight updates. Instead of costly supervised fine-tuning, FORGE converts failure experiences into natural-language memory artifacts (rules, examples, or mixed forms) and broadcasts top-performing memory instances across a population of agents between epochs, avoiding local optima. Experiments on model-agnostic benchmarks using Gemini-2.5-Flash-Lite and Llama-4-Maverick show FORGE raises average returns from -85.2 (zero-shot) to +12.6, cuts major failure rates to 1.1%, and converges 2.4x faster than single-stream Reflexion, with returns improving 1.7-7.7x over zero-shot baselines. The approach, tested mainly in the CAGE-2 network-defense environment, suggests natural-language memory density is higher than commonly assumed, though semantic saturation and applicability to fuzzy creative tasks remain open questions.

In current LLM engineering practice, there is a persistent logical gap: how to balance task specialization with computational cost. The traditional path relies on supervised fine-tuning (SFT), which comes at the price of enormous compute and potential catastrophic forgetting.

The FORGE protocol (Failure-Optimized Reflective Graduation and Evolution) offers a highly compact alternative framework whose essence is completely decoupling model intelligence from memory logic.

Mechanism Chain: From Reflection to Broadcast

FORGE's logical architecture can be abstracted as:

E_{total} = \text{Reflexion}_{inner} + \text{Broadcast}_{outer} - \text{Cost}_{grad}

> Where $E_{total}$ is total system efficiency, $\text{Reflexion}_{inner}$ is individual reflection gain, $\text{Broadcast}_{outer}$ is population broadcast gain, and $\text{Cost}_{grad}$ is the compute reduction from the graduation mechanism.

1. Atomization of Knowledge Artifacts: The system does not store raw logs; instead, reflection converts them into three artifact types:

Rules: abstracted heuristic logic.
Examples: concrete few-shot demonstrations for specific tasks.
Mixed: a synergistic combination of both.

2. Population Broadcast: This is the system's core accelerator. By distributing Top-N memory instances between epochs, the system effectively avoids the local optimum trap that traps single agents.

> Local optimum trap: an agent finds a "good enough" method during learning and clings to it, missing the chance to discover the optimal solution.

Data Density Analysis: The Logic Behind the 7.7x Return

Experimental data on Gemini-2.5-Flash-Lite and Llama-4-Maverick shows FORGE exhibits strong model-agnostic behavior:

| Metric | Zero-shot (baseline) | Reflexion (single stream) | FORGE (population) | | :--- | :--- | :--- | :--- | | Average return | -85.2 | -22.4 | +12.6 | | Major failure rate | 28% | 12% | 1.1% | | Convergence speed | N/A | 1.0x | 2.4x |

Conclusion and Limitations

FORGE demonstrates that the density ceiling of natural-language memory is far higher than currently assumed.

However, the system still faces the risk of "semantic saturation" — when the notebook of mistakes grows thick enough, do additional lessons still yield marginal utility? Moreover, experiments so far focus mainly on strongly logical game environments like CAGE-2. Whether this "failure-optimized" logic still applies to fuzzier creative or emotional tasks requires further validation.

---

Paper Details