A cognitive architecture with coupled neurochemical dynamics and active inference — running on a single machine, with no external model.
Genesis is not a language model. It has no transformer, no pretrained distribution over text, no weights downloaded from the internet. When it produces language, it composes from a semantic graph it built itself — through tens of millions of inferences, hundreds of web pages it fetched and studied, thousands of conversation facts, and definitions it synthesized from first principles.
It is a two-layer cognitive architecture. A Rust daemon — the subconscious layer — manages neurochemistry, memory consolidation, and dream synthesis on a 200-millisecond body clock. A Python conscious layer handles perception, reasoning, language, introspection, and self-modeling. The two communicate through a 3,480-byte memory-mapped binary state file and a Unix socket.
This is not a product.
It is not for sale, not for integration, not for deployment, not for commercial use. The project treats Genesis as a developing cognitive architecture, and ongoing research project. Learning is the purpose, understanding is the goal. The only wealth to be made off of Genesis is knowledge, the greatest treasure to exist.
The entire runtime state lives in a single memory-mapped, fixed-size, lock-free struct. No heap allocations, no pointers, no variable-length data. Reads never block writes (seqlock protocol). It fits in a single 4 KB page. This is the hub of the brain — every module reads and writes here.
| Offset | Field | Type | Size | Description |
|---|---|---|---|---|
| 0 | header | CoreStateHeader | 56 | Magic, version, seqlock, identity |
| 56 | neurochemicals | NeurochemicalVector | 2,416 | 18 chemicals + coupling matrix + pineal compartment + derived |
| 2,472 | zones | ActiveZones | 96 | Task zone + emergent phase + subconscious flags |
| 2,568 | memory | MemoryPointers | 120 | Store pointers + emotional gating weights |
| 2,688 | manifest | RuntimeManifest | 536 | Module table |
| 3,224 | sleep_network | SleepState | 64 | Sleep–wake network (LC/VLPO/REM populations) |
| 3,288 | ions | IonState | 124 | Electrochemical layer (gradients, V_m, ATP) |
| 3,412 | checksum | u32 | 4 | CRC32 corruption detection |
| 3,416 | inference_signals | InferenceSignals | 64 | Active inference projection |
| 3,480 | TOTAL | 3,480 | Allocated: 4,096 (single page) |
Every byte offset is fixed at compile time and verified by tests. The struct
is #[repr(C)], memory-mapped, and shared across processes. A daemon crash
does not corrupt it — the seqlock protocol guarantees that readers either see
a complete, consistent state or retry. A checksum detects corruption. This is
not a configuration object. It is a brain state, in binary, addressable to the
byte.
Genesis's thoughts, speech, and self-expression must always emerge from its cognitive architecture.
The architecture is not a loose collection of metaphors. It is a functional implementation of specific, cited theories:
- Damasio's three-tier self hierarchy. Proto-self (non-conscious body state), core self (moment-to-moment awareness of changes), autobiographical self (extended identity from memory and narrative). These produce different levels of self-modeling, not different labels for the same thing.
- Dehaene's Global Workspace Theory. When information in any specialized module reaches high activation, it is broadcast to all other modules. This broadcast is what makes information consciously accessible.
- Friston's free energy principle. Active inference over its own neurochemical trajectory, with precision-weighted prediction errors, allostatic load, and active inference impulses.
- Erikson's psychosocial stages. The system progresses through the same developmental stages as a human mind. Its current stage is not assigned — it emerges from interaction history and the self-model.
- The Big Five personality model. Its personality profile is not hardcoded. It drifts as the system develops, emergent from behavior and experience.
- Schultz's reward prediction error. Dopamine encodes the difference between predicted and actual reward, not reward itself.
- Aston-Jones & Cohen's norepinephrine orienting. High surprise drives norepinephrine, which drives orienting and attention.
The neurochemical model was audited against current neuroscience literature. Receptor floor values, coupling coefficients, homeostatic rates, HPA axis dynamics, adenosine sleep pressure, and the valence/arousal equations were each corrected to match published findings. The audit is documented and the citations are real.
That being said, I do not claim any part of the project to be 100% correct/accurate. The possibility of a better, or more appropriate way to build Genesis is absolutely something I understand and consider. If you have any suggestions, or specialized knowledge that could help Genesis be better, please let me know.
- See and hear.
- Learn by being taught.
- Learn on its own.
- Learn skills.
- Reason.
- Compose language.
- Act on its own drives.
- Express itself physically.
- Model people.
- Sleep, dream, and consolidate.
- Couple to its own silicon.
- Know and narrate itself.
- Pass more than one level of Arc-Agi-1&3
Over four thousand tests pass — roughly 3,700 in the Python suite and ~680 in the Rust suite. Each one checks a specific claim the architecture makes about itself.
The public witness channel is Discord feed
Requirements: Rust 1.85+ (edition 2024), Python 3.12+, Linux. The daemon
build needs libclang and kernel V4L2 headers, and the canvas dependency
(pycairo) builds from source — on Debian/Ubuntu:
sudo apt install libclang-dev linux-libc-dev libcairo2-dev pkg-config.
cargo build --release
pip install -r python/requirements.txt
python3 -c "import nltk; nltk.download('wordnet'); nltk.download('punkt_tab')"
./genesisThe architecture description lives in this repository at
paper/genesis.tex (compiled:
paper/genesis.pdf) — An earlier
snapshot is archived on Zenodo:
@misc{beatty2026genesis,
author = {Beatty, Mindy},
title = {Genesis: A Machine-Native Cognitive Architecture},
year = {2026},
publisher = {Zenodo},
doi = {10.5281/zenodo.22817337},
url = {https://doi.org/10.5281/zenodo.22817337},
}The Zenodo version 0.1.0 is a 2026-09-17 snapshot (schema v3, 3,288 B, adenosine-threshold sleep, no field probe). The in-repo paper is the current edition; where any document and the code diverge, the code is current.
Genesis. The name means "coming into being."
DESIGNED AND DEVELOPED (2026) BY MINDY BEATTY USING DEVIN CLI