DEV Community

Cover image for I Gave 3 AI Agents a Decaying Notepad and They Built a Culture
Shridhar Shah
Shridhar Shah

Posted on • Edited on

I Gave 3 AI Agents a Decaying Notepad and They Built a Culture

#python #ai #machinelearning #llm

One shared memory that keeps fading, three minimal agents, no boss — and a story that outlives every note that carried it.

TL;DR: Give a few bare-bones agents nothing but a shared notepad that constantly fades, and they spontaneously build a culture — a story they keep alive together, long after any single note has decayed to nothing. No coordinator, one tiny rule. Runs on a laptop.

The setup

Three agents share one whiteboard. Every tick, everything on it fades; a note left alone dies in ~4 ticks. Each tick an agent scribbles a noisy observation, and a swarm agent also reinforces whatever's currently strongest. That's the whole rule. It's stigmergy — how termites build cathedrals by reacting to each other's mud (Grassé, 1959) — and the same dynamic that lets populations of agents settle on shared conventions (Perez et al., 2024).

The 10-second version

❌ Each on its own ✅ The swarm
A shared story exists 11% of ticks 100%
Same story tick-to-tick 3% 99%

How it works

No agent has a memory of its own. The only shared state is pad — a dict of note -> strength that fades every tick. Each tick, three agents scribble a noisy observation, and the "culture" rule is a single line: reinforce whatever the group is already backing.

pad = {}                                    # the shared, fading notepad

for tick in range(100):
    for note in list(pad):                  # 1. everything decays; faint notes are forgotten
        pad[note] *= 0.5
        if pad[note] < 0.1:
            del pad[note]

    for _ in range(3):                       # 2. three agents act this tick
        obs = random.choice(FACTS)           #    each jots a weak, noisy observation
        pad[obs] = pad.get(obs, 0) + 0.3
        if imitate:                          #    the ENTIRE culture rule:
            leader = max(pad, key=pad.get)   #      reinforce what the group already backs
            pad[leader] += 1.0
Enter fullscreen mode Exit fullscreen mode

Flip imitate off and you get pure noise. Flip it on and a single story takes over — chosen by no one, kept alive by everyone.

The part that got me 

Tick 0:   group rallies around  "water down"
Tick 100: group still holds     "water down"
   ...but that original note faded to ~8e-31 within a few ticks.
   It survived only because the agents rewrote it 300 times.
Enter fullscreen mode Exit fullscreen mode

The story is 100 ticks old; the note that started it has been gone since tick ~4. What persists isn't any note — it's the meaning, re-inscribed by the group. It's the Ship of Theseus: no original plank left, yet the ship sails on. Reseed and a different story wins — the agreement is what's real.

Why it matters

The proven part: stigmergy — coordination through traces left in a shared medium — is old biology (Grassé, 1959), and populations of LLM agents have already been shown to form and transmit shared conventions (Perez et al., 2024). This demo distills it to a single rule on a decaying pad.

Where it's heading: we keep trying to fix agent memory with bigger storage. This points the other way — a group of forgetful agents in 2026 can hold knowledge no single one could, just by reminding each other. Persistence becomes a property of the society, not the context window.

These agents can't remember much alone. Together, on a whiteboard that won't stop erasing itself, they keep a story alive as long as they care to. That's culture.

How faithful is this?

This is stigmergy distilled to one reinforcement rule on a decaying dictionary — not LLM agents reasoning in language. It shows the dynamic (a shared story outliving the notes that carried it); the cited work studies it with real populations of generative agents.

Try it 

git clone https://github.com/Shridhar-2205/secret-lives-of-agents
cd secret-lives-of-agents/02-emergent-culture && python demo.py
Enter fullscreen mode Exit fullscreen mode

The series — The Secret Lives of AI Agents

  1. Agents invent their own language
  2. Agents build a culture on a decaying notepad (you're here)
  3. Agents that live inside dreamed-up worlds

Shridhar Shah — Senior Software Engineer on the AI team at Cisco. GitHub · LinkedIn

Sources & further reading: Stigmergy (Grassé, 1959) — the biology of indirect coordination · Park et al., Generative Agents: Interactive Simulacra of Human Behavior (2023) · Perez et al., Cultural Evolution in Populations of Large Language Models (2024).

Top comments (0)