ARCHIVES

The dominant paradigm for improving artificial intelligence systems over the last decade has been scale: larger models, trained on more data, with more compute, produce better performance. This paper describes a complementary and comparatively inexpensive paradigm, which we term Software Neurotechnology — a class of software layers that sit around a trained model and shape how it remembers, reasons, regulates itself, models the people and world it interacts with, and improves across repeated use, without altering the model's underlying weights. The term is a deliberate parallel to neurotechnology in the biomedical sense: just as a cochlear implant or a memory prosthetic augments a biological nervous system through an external interface rather than by rewiring the brain itself, these software layers augment an artificial one by restructuring its inputs, intermediate computation, and outputs. We present a twenty-layer framework organized into five categories — Memory, Reasoning & Planning, Self-Regulation, Interaction & World-Modeling, and Learning & Adaptation — describing the function of each layer, its relationship to existing empirical research, and its expected contribution to the practical, functional intelligence of a deployed system. We argue that this class of technique, largely orthogonal to model scale, represents an underexplored and comparatively cheap lever for improving AI systems, and we offer the framework as an organizing lens and a practical checklist for system builders.