FORGE: Self-Evolving AI Agent Memory via Population Broadcast, No Weight Updates

Forum topic · QianXun · 2026-05-19

Summary

A May 2026 arXiv paper (ID 2605.16233) by Igor Bogdanov, Chung-Horng Lung and colleagues from Carleton University introduces FORGE, a framework enabling self-evolving AI agents without any weight updates. FORGE works through two loops: an inner loop where a reflection agent distills failed task experiences into memory artifacts (rules, examples, or mixed types), and an outer loop called population broadcast, in which the best-performing elite agent's distilled memory is transmitted to the entire agent population. On the CybORG CAGE-2 cybersecurity defense benchmark, FORGE improved performance up to 7.7x over zero-shot baselines, with weaker base models benefiting most. The article also raises open questions: crude merging of multiple elite agents' knowledge, unbounded memory growth versus limited context windows, and survivorship bias from lucky high scores. FORGE suggests AI progress may come from shared 'cultural' memory rather than bigger models.

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

Source: arXiv paper 2605.16233 (May 2026), by Igor Bogdanov, Chung-Horng Lung et al. (Carleton University and collaborators). Forum post translated and summarized from zhichai.net.

The post opens with an analogy: a fishing tribe where villagers gather nightly around a fire to share lessons from failures ("avoid the eastern reef by going around to the left") dramatically raises the whole tribe's survival rate. FORGE applies this principle to AI agents — sharing distilled failure lessons across a population instead of each agent learning in isolation.

The Problem: AI's Island Dilemma

The post identifies two extremes in prior approaches:

Fine-tuning: modifying model parameters to learn new patterns — expensive, like performing surgery to teach someone to fish.
Reflection (e.g., Reflexion): an agent writes a summary after failure, but it's "island learning" — a pit that Agent A falls into will still befall Agent B.

FORGE connects these islands into a network, evolving agents with zero weight updates.

Architecture: Two Loops

1. Inner Loop — the midnight campfire review

Rules (e.g., "never execute scripts without validation")
Examples (e.g., a successful prior SQL injection defense)
Mixed (both)

2. Outer Loop — Population Broadcast

Results

On the demanding CybORG CAGE-2 cyber-defense benchmark, FORGE improved performance by up to 7.7x over the zero-shot baseline.
Notably, weaker base models benefit the most, narrowing the intelligence gap caused by model scale — like a weaker hunter becoming competent once given the tribe's best playbook.

Open Questions the Author Raises

1. Crude knowledge fusion: the system merely broadcasts the single champion's memory; how to merge breakthroughs from multiple agents excelling in different specialties remains vague in the paper. 2. Memory inflation: rule-type memory saves tokens versus examples, but the accumulated "ancestral playbook" keeps growing. When to prune stale entries within limited context windows is still an engineering black box. 3. Survivorship bias: if an agent scores high purely by luck, could "superstitious rules" be broadcast to everyone? How the system distinguishes genuine patterns from random success is unaddressed.

Takeaway

The author concludes that real wisdom lies not in never failing, but in making failures belong to everyone. FORGE suggests the future of AI may rest less on ever-larger models and more on robust cultural carriers — a self-updating, collectively shared "silicon-based culture" built through language, without touching expensive parameters.

> Truth lives forever through sharing; logic resonates through broadcast.