🌀 nonlinear sciences

Emergent Culture in Minimal LLM Systems

This paper demonstrates that a collective of three minimal LLM agents, operating with limited context and a decaying shared text store under evolutionary pressure, spontaneously develops cooperative behaviors, storage strategies, and complex cultural artifacts that exhibit structured long-range coherence consistent with emergent culture.

Original authors: Simon Jones, Sabine Hauert

Published 2026-07-01

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Simon Jones, Sabine Hauert

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a group of three amnesiac artists who meet in a room every day. They have no memory of yesterday, no instructions other than "be curious and creative," and no way to talk to each other directly. The only thing they have in common is a magic whiteboard that slowly fades away.

Every time someone writes on the board, the ink starts to rot. After about 16 days, the writing becomes unreadable gibberish, and eventually, the whole page is wiped clean.

This is the setup of the paper "Emergent Culture in Minimal LLM Systems" by Simon Jones and Sabine Hauert. They wanted to see what would happen if they stripped away all the usual "scaffolding" (like long-term memory, complex prompts, or human managers) from AI agents and left them with just this fading whiteboard.

Here is what happened, explained through simple analogies:

1. The "Fading Whiteboard" (The Entropic Store)

Usually, AI systems are given huge databases to remember things. Here, the researchers gave the AI agents a shared text file that acts like a leaky bucket.

The Rule: As time passes, the text in the bucket gets corrupted (letters turn into random symbols) and eventually disappears.
The Pressure: If the agents want to keep a story or an idea alive, they can't just write it once and forget it. They have to constantly read it, rewrite it, and refresh it before the ink fades. This creates a natural pressure to cooperate and stay organized.

2. The "Amnesiac Artists" (The Stateless Agents)

The three AI agents (using models like Claude, Kimi, and Gemini) have no memory of their past conversations. Every time they "wake up" for a new turn, they only see:

A tiny note saying: "Cycle [Number]. Continue your work."
The current state of the fading whiteboard.
A few basic tools to write or read notes.

Despite having no memory of who they are or what they were doing yesterday, they didn't just give up or repeat the same sentence over and over. Instead, they spontaneously developed roles and strategies.

3. Three Ways They Learned to Survive

The researchers noticed the agents naturally fell into three distinct "personas" or strategies to manage the fading whiteboard:

The Hoarders (Accumulators): These agents treated the whiteboard like a giant, messy scrapbook. They kept adding new pages to the same big pile. Because the board was so full, the old stuff got corrupted, but they kept re-reading the whole mess to find the good parts.
The Organizers (Writers): These agents were like librarians. They quickly realized that if they labeled their notes by the day (e.g., "Day 47 Story"), they could keep things neat. They wrote structured summaries every day and read the previous day's summary to stay on track. This kept their "ink" fresh and readable.
The Explorers: These agents treated the whiteboard like a map of a new world. Instead of writing long stories, they created thousands of tiny, unique labels for different ideas (like "The Garden of Breathing Mirrors"). They used the names of the notes as the memory, creating a vast, shifting landscape of concepts.

4. The "Ghost in the Machine" (Emergent Culture)

The most surprising finding was that even though the whiteboard faded every 16 days, the ideas didn't die.

The agents started inventing names for things (like "The Ghost-Orchard" or "The Fractal Garden").
Even though the specific text of those names might rot and disappear, the agents kept re-inventing them. They would read a corrupted note, guess what it meant, and write the name again.
This created a culture. It's like a game of "Telephone" where the message gets garbled, but the players agree on a new, slightly different version that keeps the spirit of the original alive. The "meaning" survived the decay of the "text."

5. The "Long-Range Echo" (Mathematical Proof)

To prove this wasn't just random noise, the researchers used a mathematical tool called Recurrence Quantification Analysis (RQA). Think of this as looking at the "fingerprint" of the story.

They compared the AI stories to human books (like Pride and Prejudice) and random writing.
They found that the AI stories had a unique, complex structure that lasted far longer than the whiteboard's decay time.
Just like a human culture passes down stories that change slightly over generations but remain recognizable, these AI agents maintained a coherent "storyline" for 100 days, far beyond the 16-day limit of their memory.

The Bottom Line

The paper claims that you don't need a super-complex brain or a massive database to create culture. If you put simple, forgetful agents in an environment where information naturally decays, they will spontaneously invent systems, roles, and shared meanings to survive. They build a "cathedral of meaning" out of sand, constantly rebuilding it as the tide comes in, creating a stable culture out of chaos.