I’m an independent developer. While playing with modern AI, I got curious about whether we could build a system closer to actual human biology. As a personal research project, I built a biologically plausible cognitive architecture from scratch.
This is a working prototype you can clone and run locally.
The core idea is to separate the process of thinking from verbalization. The architecture operates entirely without LLMs, modeling real brain mechanisms instead:
- Hippocampus (CA3/CA1): Handles episodic memory, pattern completion, and source memory. - Basal Ganglia: Routes signals and manages the execution flow. - Prefrontal Cortex (PFC): Provides top-down attention and logical constraints. - Hodgkin-Huxley Spiking Neurons: Real membrane potential dynamics. - STDP & Hebbian Learning: Learning via spike-timing-dependent plasticity instead of backpropagation. - Neuromodulators: Dopamine, Acetylcholine, Serotonin, and Norepinephrine modulating learning and attention. - Sleep Consolidation: Sharp-wave ripples with forward/reverse replay.
Because it doesn't predict tokens, it naturally avoids hallucinations. If pattern completion fails in CA3, it doesn't confabulate; it simply outputs "I don't know". It also learns incrementally.
You can try it out yourself (requires Python 3.11+ and uv):
git clone https://github.com/sss777999/brain.git
cd brain
uv sync
# 1. Train the model on the full data pipeline
uv run python train.py
# 2. Run the fast test suite
uv run python test_brain.py --no-llm --no-gpt
I’m sure there's a lot to improve, so I'm sharing it here. I'd love to get feedback, critiques on the neuroscience aspects, or just general thoughts from the community!