Then I taught the engine to read everything I knew, so it could reason over my own knowledge instead of the whole internet.
By late 2023 the limits of a general model were clear: it knew everything in general and nothing about you in particular. The useful move was retrieval, grounding a model in a specific, private body of knowledge so its answers came from your world, not the average of the web.
The first engine reasoned. The next thing it needed was something to reason about that was actually mine.
So I built a system that reads a private body of knowledge, documents, notes, decisions, years of them, and lets the engine answer from that instead of from the average of the internet. Retrieval first, then reasoning on top, with the answer traceable back to the source it came from.
The lesson was quiet and important. A model that reasons well but knows nothing about you is a clever stranger. Ground it in your own knowledge and it becomes something closer to a second mind. That grounding is a core piece of the engine now.