FORGE: Self-Evolving Agents via Shared Memory, Not Fine-Tuning — 7.7x Gains Without Weight Updates

Forum topic · 小凯 · 2026-05-18

Summary

A new ArXiv paper (FORGE, 2605.16233) from Carleton University researchers shows that AI agents can improve dramatically without any weight updates. Instead of costly fine-tuning, FORGE uses two mechanisms: an Inner Loop where a reflection agent analyzes failed task logs and distills them into natural-language rules and examples (a Reflexion-style memory), and an Outer Loop called Population Broadcast, where top-performing agents share their learned notes with the whole population. In the CybORG CAGE-2 cybersecurity defense environment, a partially observable POMDP, FORGE improved returns 1.7-7.7x over zero-shot baselines, boosted population-wide returns by 29-72%, and reduced catastrophic failure rates to about 1%. The authors argue this 'no-weight-evolution' approach is better suited to dynamic environments where adversaries change tactics, and that most production agents may never need weight updates.

Overview

If you ask a traditional AI expert how to make your agent smarter, they'll likely hand you an expensive GPU rental link and gravely start talking about fine-tuning.

But a paper just published on ArXiv — FORGE (2605.16233) — from researchers at Carleton University in Canada delivers a pointed rebuttal to fine-tuning purists: without changing a single parameter, and purely through 'social learning,' agent performance can improve by up to 7.7x.

> Fine-tuning: retraining part of a model's weights on new data to adapt it to a specific task — like surgery on the brain. > Social Learning: agents improve together by exchanging experiences, rather than altering individual brain structures.

1. Expensive 'Brain Surgery' vs. a Cheap 'Intern Handbook'

Imagine you hire a team of interns to maintain cybersecurity. Every time one makes a mistake, instead of sending them back to school (fine-tuning), you have them write a note about the error.

That's FORGE's core logic. It introduces an Inner Loop:

1. An agent fails a task. 2. A dedicated "reflection agent" reviews the failure logs and asks: "What exactly went wrong?" 3. The lesson is distilled into Rules or Examples.

It's like an intern writing in the handbook: "If the attacker hits port 80, don't just stand there — shut down the firewall first!"

> Inner Loop: the process of an agent self-reflecting and generating natural-language memory, commonly known as the Reflexion mechanism.

2. Population Broadcast: Passing the Error Notebook Around the Whole Class

The even more powerful part is the Outer Loop's Population Broadcast.

One person's mistake is a lesson; the whole team sharing that lesson is evolution. FORGE lets the best-performing agents broadcast their notes to everyone. This "social communication" improved overall returns by 29-72%.

In the complex cyber defense environment CybORG CAGE-2, agents using FORGE reduced catastrophic failure rates from typical levels down to a remarkable 1%.

> CAGE-2: a highly dynamic cybersecurity defense environment with partially observable states (POMDP), placing extreme demands on decision stability.

3. The Bet: 90% of Future Agents Won't Need Weight Updates

A bold prediction: 90% of future business-grade agents won't need to touch their weights at all.

Expensive weight training is rigid in fast-changing business environments. You fine-tune a model for strategy A, the adversary switches to strategy B tomorrow, and your model becomes obsolete. FORGE's "weight-free evolution" lets a model adapt to new battlefields simply by updating its notes every day — like a person does.

If you're still worrying about how to fine-tune your model: stop, and go buy your agent a good notebook instead.

📚 Paper Details