What it does
Five networks, perception, memory, emotional processing, reasoning and communication, interact over a lifespan rather than a training run. Memory is written to files and inherited across generations; a dream generator consolidates during idle periods; lifespan parameters are randomized so no two runs are the same shape.
It is the conceptual seed the rest of the lab grew from: AXON, MIRROR and Archie are all descendants.
In the code
# Emotional memory persists to disk between runs and is inherited by the
# next generation. Keeping it as plain files rather than model weights
# means it can be read, diffed, and reasoned about by a person.
memory_dir/
├── episodic/ # what happened, timestamped
├── emotional/ # valence and intensity attached to episodes
└── inherited/ # carried forward from the previous generation
A developmental system whose memory you cannot read is a black box with extra steps. Files make the development observable.
How this differs from the ordinary version
Learning from experience, not corpora
The whole point is what emerges from interaction over time rather than from bulk pretraining. Runs are not reproducible in the usual sense, which is a property of the object of study, not a defect.
Generational inheritance
Memory carries across lifespans, so behaviors accumulate over generations. This is the mechanism that made the developmental questions worth pursuing into AXON.
In the field
What it is for
This is research, not a product, and it is listed as research. Its value to the rest of the work is that the persistence and self-observation ideas that make AXON auditable were first tested here where the stakes were zero.
Questions
- Is this AGI research?
- No. It is a small developmental architecture for studying how persistent memory and emotional weighting change behavior over time.