$ cd /projects/caveman
2026-02-24

← back

Caveman — Survival-Driven AI Agents

GitHub: private


PRE — Idea · Setup · Build

Goal: What happens when you give simple entities survival constraints,
free will, and 64 memory slots? Do they behave like... humans?

I always wanted to see how real intelligence would behave given certain
constraints. Basically a simulator like Conway's Game of Life — which
I discovered AFTER I built my own version. Classic.

I thought I'd need a neural net for emergent behaviour. Turns out: a
few hundred lines of Python and some Claude-generated logic are enough.
The agent ("Grug") has exactly 3 drives: hunger, pain, fatigue. A
64-slot memory that forces importance-based forgetting. And a simple
perception → recall → decide → act → update loop.

That's it. No transformers. No embeddings. No GPU. Just constraints.

Stack:
  - Python (PyGame for simulation/visualization)
  - Claude (code generation, logic design)
  - No external APIs at runtime — pure local execution

Architecture:
  - Entities with running-average memory (not exact facts)
  - Worlds with food, dangers, physics
  - Single agent ("Grug") or tribe mode (multiple agents)
  - Hyperspeed mode for batch simulation runs

Success metric: Did Caveman survive longer this run than last? That's it.

POST — Learnings · Afterthoughts · Timeline

Learnings:
  - You don't need neural nets for emergent behaviour. Constraints
    plus simple rules plus memory limitations produce surprisingly
    human-like patterns. Grug hoards food. Grug avoids pain. Grug
    forgets unimportant things. Sound familiar?
  - The results of this experiment are basically exactly what I
    already knew from 50 years of watching humans. Intelligence
    isn't about complexity — it's about survival under constraints.
  - 64 memory slots forces prioritization. The agent MUST decide
    what's worth remembering. That's not a limitation — that's
    what makes it realistic. Humans do the same thing.
  - This was a philosophical precursor to Friend's ANIMA system.
    Same idea (drives that decay, constrained resources) but
    Caveman is the minimal proof that the concept works.

Timeline:
  - 2024-12: Built in a few days. Experimented. Watched Grug
    survive (or not). Nodded knowingly. Moved on.

Status: Complete. The experiment proved what it needed to prove.
  Sometimes the best projects are the ones you finish quickly.