A geometry built
to think.

Before describing what PCM does, it is worth saying who else builds this way. The answer is: every mind we have ever observed in nature.

The brain does not store its knowledge in weights. It stores it as structure. The cortex is full of topographic maps that preserve geometry from one layer to the next. The primary visual cortex preserves the spatial arrangement of the retina; the somatosensory cortex preserves the topology of the body; the motor cortex preserves the layout of the muscles it controls. These are not pedagogical simplifications. Adjacent neurons in V1 actually respond to adjacent regions of the visual field. The shape of the cortex carries information that the firing rates by themselves do not.

The hippocampus is more striking still. Place cells fire when an animal occupies a particular location in its environment. Grid cells fire at the vertices of a triangular lattice that tiles whatever space the animal moves through. Together they form a coordinate system, a literal geometric representation of space, that the rest of the brain reads. The 2014 Nobel Prize in Physiology recognized this work because it changed how neuroscience thinks about representation itself.

In the last decade, the picture has generalized. The activity of large populations of neurons in many cortical areas traces out continuous low-dimensional structures, what the field now calls neural manifolds. The manifold is where the meaning lives. Two motor neurons can fire in different ways across two reaching motions, but the population trajectory on the manifold tells you which motion was being performed. The geometry, not the cells, is the signal.

We believe this is not coincidence. Geometry is what intelligence runs on, in the only working examples of intelligence we have. PCM is built on the same premise, applied to artificial substrate.

Knowledge is not a pile of weights. It is a structure with shape.

A conventional neural network represents what it knows in a matrix of weights, set once during training. Adding new information requires recomputing that matrix, which is why today's systems are frozen between training runs. They cannot integrate new experience while they operate; they can only generate against what they were given.

PCM takes a different primitive. Knowledge is stored as curvature on an internal manifold, a continuous medium that admits local updates. Adding information bends the surface in one region without disturbing the rest. Recall reads the shape. Learning and inference are the same kind of operation performed on the same medium, at the same time.

We believe this is the difference between a model and a mind. A model represents a snapshot of the world. A mind keeps a structure that the world can keep changing.

Operation and learning are the same activity, run on the same medium, at the same time.

A PCM in service does not pause to learn. There is no offline window, no retraining cycle, no fine-tuning step. Each input nudges the underlying geometry. Most of those nudges dissipate. Some are committed to durable structure, and from that moment on the system can recall them cheaply. By the time the system has answered a question, it is already a slightly different system than the one that received it.

We believe this is closer to how cognition actually works. The brain does not retrain itself at night and then resume operating in the morning. It learns while it lives.

Compute cost grows with the logarithm of what is known, not linearly with the size of the system.

A conventional model becomes slower and more expensive in direct proportion to the volume of knowledge it has been trained on. Doubling its capability roughly doubles its compute. PCM scales differently. Because settled knowledge is organized for cheap retrieval, the cost of running a PCM tracks the logarithm of its accumulated knowledge rather than the size of it.

The downstream consequence is operational: a PCM gets faster, not slower, as it learns the domain in which it operates. Power draw follows a similar curve. We believe this is what makes a continuously learning system actually deployable in edge and embedded settings, where compute and power are constrained.

During quiet periods, the system reorganizes its own interior.

When external demand is low, a PCM continues to work. It reconsiders structures still under evaluation. It looks for inconsistencies across regions of settled knowledge that have grown in parallel without ever being compared. It consolidates, prunes, and tightens what it has accumulated.

We use the word dreaming deliberately. It is not a marketing flourish. The process is structurally analogous to the consolidation that happens during sleep in biological cognition: the system reviews recent experience, decides what is worth keeping in durable form, and integrates it with what was already known. A PCM that has been running quietly overnight is meaningfully better in the morning than the PCM that finished its last task.

The architecture is not a research thesis. It produces operational properties that systems built on frozen weights cannot match.