Patina

Problem

Every AI conversation starts from zero. The system forgets what you told it yesterday, repeats mistakes you already corrected, and never develops judgment about how you work. I built Patina because I wanted an AI that accumulates understanding: a persistent belief graph of user judgment that compounds over time instead of resetting.

Architecture

By the numbers

Self-improvement thesis

The interesting constraint: how do you let a system improve itself without losing control? Patina tracks which beliefs lead to good outcomes and which get corrected. It proposes changes to its own behavior through the same PR pipeline it uses for everything else, so no self-modification ships without human review. The system gets better, stays safe, and remains interpretable.

Novel contribution

Most AI agents either do too much unsupervised or too little to be useful. Patina instantiates CAIS principles in a personal cognitive system: the observer and builder are separated by an air gap, the human merge gate bounds autonomy without crippling capability, and the self-improvement loop is governed, not open-ended. It's graduated autonomy as a running system, not a paper.