Persona: A Growable Personality Layer That Turns Any Agent Into a Companion That Learns You

Today's AI has intelligence but no experience. Every session starts from zero, so agents repeat the mistakes they were corrected for, deliver the same generic work to everyone, and pay again to rediscover what they already found. Fine-tuning does not fix this: it forgets, costs a training run, hides the lesson in weights, and is lost when the model is replaced. We present Persona, a personality layer that separates experience from intelligence. A pblock is a persona's experience as plain data: real situations with their confirmed outcomes and small calibrated judgment heads over a frozen 22 MB embedder. It is trained before use on a consumer CPU, composed into panels for the work at hand, keeps learning from confirmed outcomes, is corrected without touching the model, and moves unchanged to a new model, machine, or team. Because every confirmed outcome becomes a lesson, a Persona is designed to compound with use. Before each task the panel tells the model what worked and what failed, citing the cases behind it. The same engine is a guard: Keel Guard is an antivirus for agent requests, updated by training new cases rather than retraining a model. In first experiments on one consumer PC, the engine grew personas in eleven domains; raised a host agent from 83.0% to 85.4% in a sealed five-field test; on SWE-bench Verified held accuracy (89.0% vs. 88.0% over 1,000 paired runs) with fewer steps and ~12% less compute; and, as Keel Guard, stopped 89.4% of never-seen harmful requests at 34 ms on a CPU, with its 55% false-alarm rate as the next training target. We release the engine, the trained stores, and the full evidence.