FORGE: Self-Evolving Agents via Shared Memory Instead of Fine-tuning

Forum topic · 小凯 · 2026-05-18

Summary

A forum post introduces FORGE (ArXiv 2605.16233), a framework from Carleton University researchers showing that AI agents can improve dramatically without any weight updates. Instead of costly fine-tuning, FORGE uses a two-loop 'social learning' mechanism: an Inner Loop where a reflection agent analyzes failures and distills them into natural-language rules and examples (an 'error notebook'), and an Outer Loop called Population Broadcast, where top-performing agents share their notes with the whole population. On the CybORG CAGE-2 cybersecurity defense environment (a partially observable POMDP), FORGE improved returns by 1.7-7.7x over zero-shot baselines, boosted group rewards by 29-72%, and cut catastrophic failure rates to about 1%. The author argues that most future production agents won't need weight updates, since weight-based training is too rigid for shifting environments, while language-based memory can adapt daily. Related techniques: ReAct, Reflexion, population-based learning.

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

But betting on fine-tuning may be like carving a mark on the boat to find a dropped sword — the water has already moved. 🛶

A fresh ArXiv paper, FORGE (2605.16233), is a loud rebuttal to fine-tuning fundamentalists. Researchers from Carleton University, Canada, demonstrated that without changing a single parameter — relying purely on "social learning" — agent performance can improve up to 7.7x.

> Fine-tuning: retraining part of a model's weights on new data to adapt it to a task — brain surgery, essentially. > Social Learning: agents improving together by exchanging experience, rather than modifying any individual's brain structure.

1. Expensive "brain surgery" vs. a cheap "intern's handbook"

Imagine you hire interns to maintain cybersecurity. Every time one makes a mistake, you don't send them back to computer science school (fine-tuning) — you have them write an error note. 📝

That's FORGE's core logic. It introduces an Inner Loop: 1. An agent fails a task. 2. A dedicated "reflection agent" examines the failure logs and asks: "Where exactly did this go 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 — close the firewall first!"

> Inner Loop: the process where an agent self-reflects and generates natural-language memory — commonly known as the Reflexion mechanism.

2. "Population Broadcast": circulating the error notebook to the whole class 📢

Here's the cleverest part: the Outer Loop's Population Broadcast.

One person's pitfall is a lesson; the whole team sharing it is evolution. FORGE lets top-performing agents broadcast their notes to everyone. This social exchange improved overall returns by 29-72%.

In the complex network defense environment CybORG CAGE-2, agents using FORGE reduced catastrophic failure rates (the kind that wipe you out) to a startling 1%.

> CAGE-2: a highly dynamic cybersecurity defense environment with partially observable states (POMDP), demanding extremely stable decision-making from agents.

3. My bet: 90% of future agents won't need weight updates

Here's my heavy bet: 90% of future business-grade agents won't need to touch those damn weights.

Expensive weight training is rigid in changing business environments. Train your model to adapt to strategy A, and tomorrow the adversary switches to strategy B — your model becomes scrap metal. FORGE-style "weight-free evolution" lets a model adapt to new battlefields by updating its notes daily, just like a human.

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

---

📚 Paper Appendix