A personal research experiment · Solo research and engineering
Prometheus: a real-time cognitive architecture with no LLM
Build, from scratch and without any LLM, an architecture where an embodied agent can compute about its own internal states in real time.
- stack size
- ~114,500 parameters, multi-layer
- loop latency
- 2.8 ms per loop (real-time)
- death proximity accuracy
- 97.7% (from 4 positive examples)
- death signal jump
- 5.2x (0.159 to 0.832)
- Embodied agent
- 8-D continuous affect space
- Layered self-model
- Real-time loop
Problem
Prometheus is a personal research experiment that tests a build question: can an artificial cognitive architecture be stood up from scratch, without leaning on an LLM at all? The brain's modules, attention, memory, threat, self-model, decision, are coded separately, and an embodied agent in a virtual world has to compute about its own internal states.
Constraint
The hardest constraint turned out to be operational, not philosophical: embodiment.
Roughly 90% of development time went into keeping the agent alive in its physical world, falling into pits, eating mechanics that would not fire, mining the wrong block. Before any "cognition" layer could matter, the agent simply had to stop dying. On top of that, the entire cognitive stack had to run in real time, milliseconds per loop, not seconds.
Architecture decision
I modeled affect as an 8-dimensional continuous coordinate space rather than a categorical label like happy or sad, and derived intent from the movement vector through that space. Self-awareness was built in layers on top: an immediate meta-cognition layer, a behavior-mode layer, and a slower higher-order evaluation layer.
The layers are wired so that a single feedback loop injects into the next decision, which keeps the system real-time and avoids a causal loop inside a single tick. The design constraint was that reflection had to cost a bounded amount of time per loop, so introspection could never stall the agent that introspection was supposed to keep alive.
Result
The full stack, about 114,500 parameters across multiple layers, runs at 2.8 ms per loop, comfortably real-time.
After calibration, death-proximity detection reached 97.7% accuracy from only four positive examples, and the mismatch signal at the moment of death rose to 5.2x its baseline, from 0.159 to 0.832. The honest center of this project is that most of the work was embodiment engineering: a self-model is only worth building once the body it models can survive long enough to be modeled.