Learning note · About 8 min read · 2026.11.02
Finishing the node is not the same as understanding
A finished tutorial, a checklist that advanced one square, a skill graph that revealed the next layer: all three create a feeling of progress. None of them require you to say why. The next bug that looks adjacent, and is not, drops that feeling on the floor.
The size of the graph adds its own weight. A chart that tries to draw every part of a craft invites you to scan the whole thing before you are allowed to work. Experienced people usually know that having heard of everything and being able to use it are different stocks. Beginners are more likely to read the chart as an obligation. The order looks like truth. It is often just a drawing.
A long AI chat loses the thread in a different way. You open it for a specific failure. Ten minutes later you are in the history of a neighboring library, a flag you will not use today, and a suggestion to rewrite the layer while you are here. The original failure is still open. It is no longer the subject of the conversation.
A Learning Tree: anchor, drill, return
The Learning Tree is how this kind of study works in HaroCue: the deepen-while-learning pattern. The name is the shape. It is not a skill chart you are supposed to fill.
You pick an anchor question first. The anchor is the sentence you actually need answered today, in your words, small enough for one work session. Then you drill one branch: a word you lack a name for, a call you have not read, an invariant you cannot pretend to know. While you drill, you keep the understanding in your words, with a source. You treat it as mastered when you can confirm it. If you cannot confirm it, it stays an inference, or an open question.
Then you return to the main path. The branch stays attached to the anchor. What you bring back is an explanation you can open again. A tick is a different object, and so is a scroll position in a chat. The main question is still the main question. Next time you start from that explanation, not from the whole transcript.
This uses the same memory rule the product already has. Automatic observations are labelled as assistant inferences. After you read the original, you can confirm, correct, or forget. Ordinary Cue uses explicitly confirmed facts. Mastered on the Learning Tree is that confirmation. It is not a stamp the model applies because the chat felt complete.
Linear chats and paths you can tick
ChatGPT Study Mode fits one guided pass. Prompts, a scaffold, and a knowledge check can slow a sitting down in a useful way. The thread itself is linear. A useful branch and a detour are written into the same transcript. Nothing marks where the branch ended and the main question still is. That comparison has its own note.
A path you can tick has a different definition of done. The node moved, so the chart treats you as having moved. People spend years in those lists, including the career maps often called roadmaps, and the squares they tick do not become understanding they can carry. Tutorial hell is the other edge of the same gap: you can follow the steps, and you still hesitate to change them, because the reason was never kept. The Learning Tree wants that reason, tied back to the problem you are in. It does not hand you a larger chart.
Why flashcards do not hold a coding sequence
Spaced repetition is strong at atomic facts. A flag, a port, a name: those make honest cards. What you actually lose in programming is often a series. This check passed, then that claim was read, then this middleware rejected the request. People who use Anki say the distinction plainly. Cards fit the edge of a concept. They fit a sequential explanation poorly.
HaroCue is not a flashcard app, and it does not schedule reviews for you. The Learning Tree covers a different job: a study session with branches, plus evidence you can retrieve. What you want later is how you understood this rejection, not a pile of due cards. Things you use often will stick from the work itself. The explanation you use rarely, and will trip on again, needs to stay attached to the original question.
Working memory collapses when every thread is open
A real study hour often has all of these open: the editor, the failing request, a tutorial tab, a skill chart, and an AI chat that has already drifted. Each one claims to be the context. Working memory cannot hold them. You either follow the last suggestion, or you tick a square that looks related. Both feel like motion. The anchor is no longer in your hand.
A tree shrinks what you have to hold. The main question is one sentence. The branch you are in is one sentence. The evidence you will bring back is one sentence, with a source. The other tabs can close. This does not make the subject smaller. It refuses to let five places all be the current task.
One bug, one branch, one return
The anchor can be a sentence like this: why does my auth middleware reject this token. You do not need the history of authentication. You need this rejection.
The branch might be JWT claims. You stop long enough to see what aud refers to on this request, what the middleware compares it with, and whether the clock or the audience is even the failure. The questioning can happen in Study Mode, or in whatever chat you already have open. What HaroCue keeps is the sentence you bring back, not a transcript of the whole detour.
The note you return with can look like this, in your words, pointing at the source you checked: this rejection happens because aud is the public client id, not the API audience. I compared this middleware with this token. Do not mark that sentence mastered until you can point at the code and the token and say it. Once you can, confirm it. The next time the failure shows up, you read that sentence first, then decide whether a different branch is the one to drill, such as expiry or the signing key. The original question is still on the tree. It did not turn into a chat about the encyclopedia of JWT.
What a later decision can cite
The note you bring back is also what a later decision can point to. When you keep a check, drop a dependency, or change an API, the reason belongs next to the choice. Remembering that you once looked at auth is too thin to support the next tradeoff.
A decision log is the natural next note in this series: after a deep dive, how your future self can query why a choice was made. This article stops at the learning thread. The return has to leave evidence a future decision can cite. What HaroCue offers today is local memory with sources, and facts you confirm. There is not yet a separate decision-log article, and this page does not treat that name as a delivered screen.
Keep the main thread
Start from the problem you are fixing, not from the whole chart. Drill one branch. Write the explanation in your words, with a source. Confirm it when you can point at it and say it. Then return to the main question. The product overview states today’s boundaries. The download page states Mac availability. Local-first means the memory lives primarily on your machine. With a cloud model enabled, relevant text can still be sent. That matches what the overview already says.
ChatGPT Study Mode vs local deep learning
An AI tutor that remembers what you mastered
Claude for students vs local deep learning