A mycelial world in 190 kilobytes
The Wisdomkeeper generates its hero, world, music and sound in code. Underneath, a 70-weight network reads how you play and reshapes the next region.
1 Nov 2026 · 2 min readThe Wisdomkeeper is an over-the-shoulder action RPG set in fractal worlds. You are a bioluminescent seed-spirit rooted in the soil of code, carrying a story home. Play it in the browser.
Nothing to download
The browser edition loads no 3D asset files. The hero, enemies, world, music and sound effects are generated in code, about 190 KB gzipped.
- Hero: lathe geometry, layered armour plates, a thruster pack.
- Look: shaders compute rim light, vein networks and bioluminescence at runtime.
- World: recursive fractal trees and rivers of light from instanced meshes.
Three ways to move
| mode | how it behaves |
|---|---|
| Ground | walk, dash, slopes limited to about 45° |
| Free flight | six axes; look to steer, Space up, C down, Shift to boost |
| Levitation | step onto a river of light and the current carries you |
The mycelial mind
Under the game sits a deliberately tiny neural network: 5 inputs → 8 hidden → 4 outputs, about 70 weights, seeded per run.
inputs aggression · accuracy · roaming · evasion · … (how you actually play)
outputs a latent world code
effect terrain ruggedness, enemy behaviour, drops, corruption in the next region
Its output is blended with a hand-written heuristic, 60% rules and 40% network, so the world adapts without ever becoming unplayable. The same principle as the desktop edition: when the model is unsure, the game’s own rules take the turn.
A choice with weight
Each region holds Remembrances, corrupted story nodes. Clearing one forces a choice:
- Purify: heal, grow an echo-tree, push back the corruption.
- Embrace the Null: take raw power and lose health. Embrace enough of them and the story ends at The Hollow Crown.
Grounded in real fungal research
The premise, that fungal networks process information, comes from published biocomputing work. Mycelium produces action-potential-like spikes. Models of those networks implement logic gates. 1 Dried mycelium behaves as a memristor usable for reservoir computing. 2 3 Its noisy spiking has even been tested as a source of true random numbers 4. The game does not simulate any of this. It borrows the idea that a network without a brain can still decide where to grow.
The desktop edition
Infinite Evolutions 0.1.2, for macOS and Windows, adds a local mind: two small models that play the Warden and the forest, about 1.3 GB, downloaded only after you agree. No account, no telemetry. Both models run on an 8 GB Windows laptop with integrated graphics, at the Low preset.
- A. Adamatzky, P. Ayres, A. E. Beasley, N. Roberts, H. A. B. Wösten. Logics in Fungal Mycelium Networks. Logica Universalis 16, 655–669 (2022)originalcorpus: ~/infinite-worlds
- J. LaRocco, Q. Tahmina, R. Petreaca, J. Simonis, J. Hill. Sustainable memristors from shiitake mycelium for high-frequency bioelectronics. PLoS One 20(10): e0328965 (2025)originalcorpus: ~/infinite-worlds
- I. Tompris, I. K. Chatzipaschalis, T. P. Chatzinikolaou, et al., A. Adamatzky, P. Ayres, G. Ch. Sirakoulis. Mycelium as a computational medium: a framework for growth modeling towards reservoir computing. Natural Computing (2025)originalcorpus: ~/infinite-worlds
- V. Stojcheva. Mycelium as a Source of True Randomness: Exploring Fungal Networks for Organic Random Number Generation. Bachelor's thesis, LUT Universityoriginalcorpus: ~/infinite-worlds
The measurements in this article come from these records in the running system.
github.com/misomalcek/the-wisdomkeeper/README.md— Wisdomkeeper README, browser and desktop editionsthe-wisdomkeeper/src/systems/mind.ts— MycelialMind, 5→8→4 feed-forward network~/content/alteranai-content/01-Wisdomkeeper Physics, Phyto-Mycelial Intelligence, and Infinite Fractal Worlds.md— Wisdomkeeper physics and phyto-mycelial intelligence
A blueprint reported a fractal-orbit optimizer reaching loss 0.0176 on two-spirals. We ran its code. The fractal term never switched on: in 36,000 steps the measured dimension never crossed the threshold that activates it.
If the loss landscape is self-similar across scales, a fixed-window optimiser is the wrong instrument by construction. A question, written down before the work, with what it would take to become a claim.
We clustered 1,046 knowledge-graph entities and found nothing, at every k, with two methods. Two clean hypotheses for why were falsified. The structure was there all along, in a different representation.