How to build a personal AI learning system in 2026

Build a Personal AI Learning System That Sticks

Treat learning as a pipeline, not a chat window. Capture sources in NotebookLM (grounded by Gemini 3), use Claude as a Socratic tutor that quizzes rather than lectures, and route the hard parts into Anki's FSRS scheduler. The system works because it forces retrieval and spacing, the two mechanisms with the strongest evidence.

Most people who ask me about an AI learning system are really asking for a better content firehose, and that is exactly the wrong instinct. More inputs do not produce more knowledge; the constraint was never access to information. I have thirteen years building AI systems and a stretch doing research at Kyoto University, and the version of "learning fast" that actually held up looks less like consuming and more like a pipeline with deliberate friction in the right places. The decades of evidence point one way: Cepeda's 2006 meta-analysis of 184 studies found that spacing practice out beats cramming it, and Karpicke's work in Science showed that retrieving knowledge builds it better than rereading. A good system is just plumbing that forces those two behaviours.

I build mine in three layers. The first is grounded capture, and in 2026 that is NotebookLM. Running on Gemini 3, its free tier holds a hundred notebooks with up to fifty sources each and half a million words per source, so you can drop in papers, transcripts, your own notes, and a book, then ask questions answered only from those sources with citations you can check. This solves the single biggest problem with using a raw chatbot to learn: the model stops inventing and starts pointing at the page. Its Audio Overviews, three a day on the free plan, turn a corpus into a discussion you can listen to on a walk, which is useful for first exposure but not, on its own, for retention.

The second layer is an active tutor, and that is where Claude earns its place. The mistake is asking it to explain things to you, because being explained to feels like progress and rarely is. Instead I set up a project with my actual sources and a standing instruction: act as a Socratic tutor, ask me one question at a time, make me reason aloud, correct me, and never hand over the answer before I have attempted it. This is John Whitmore's GROW idea from Coaching for Performance applied to a machine; the value of a good coach is the quality of their questions, not their answers, and a model instructed to ask rather than tell becomes a tireless version of that. It is also the cyborg pattern Ethan Mollick describes in Co-Intelligence, where the tool is woven into the thinking instead of replacing it.

The third layer is durable memory, and it has to be deliberate because nothing else makes knowledge survive the month. Whatever genuinely matters, I move into Anki running the FSRS scheduler, which models my personal forgetting curve and schedules each item for the moment before I would lose it, cutting review load by roughly a quarter against the old algorithm. For ideas that need to connect rather than be memorised, a linked notebook such as Reflect, at ten dollars a month with end-to-end encryption, or Mem's newer "thought partner" build keeps notes associated so older thinking resurfaces when it is relevant. The point of three layers is that each does one job well; collapsing them into a single chat window is why most people's "AI learning" produces a transcript and no retention.

What ties this to performance rather than trivia is selection, and here I lean on Naval Ravikant's idea of specific knowledge in the Almanack: the learning that compounds is the kind that is hard to train for and uniquely yours, not the generic course everyone takes. AI makes generic knowledge nearly free, which paradoxically raises the value of the idiosyncratic, judgement-heavy material you build by wrestling with primary sources. Point the system at that, not at whatever is trending. I have written more about why exploration beats optimising a known path in a separate essay.

The honest limits matter. None of this replaces thinking; it removes the friction that kills systems by the third week, which is the real reason people quit. A model will also confidently mislead you, so the grounded-sources layer is not optional hygiene, it is the difference between learning and absorbing plausible fiction. And the scheduling tools only work on knowledge you have already made discrete and well-formed; they cannot rescue vague notes. So start small and concrete: pick one domain that genuinely advances your work, ground it in NotebookLM, let Claude quiz you until you stumble, and send only the gaps to Anki. A narrow system you run daily beats an elaborate one you admire and abandon.