Learning note · About 6 min read · 2026.11.02
Homework bots are not the tutor you need
Search for an AI tutor and the page fills with homework. Math solvers, essay drafters, assignment checkers. Someone who builds needs a different object: a tutor that already knows what you can explain, because you confirmed it, on a machine you control.
Remembering is too small a word if it only means the next chat is more polite. You want the conclusion that already held last week to still be there, and you want to open it, check it, and change it when it is wrong. An impression you cannot open is only a feeling the model has about you.
Chats forget, and checklists only move
A chat tutor forgets in practice even when the transcript is saved. Next session you restate the project, the constraint, and the thing you settled on Tuesday. A checklist app fails from the other side: it marks done when you click, and no understanding is attached. Both send you back to the same explanation. The cost is not only time. You stop trusting your own sense of what you know.
The retelling also repeats across tools. The morning conclusion lives in one chat. The afternoon editor is another model. The evening review is a third. Each one meets the repository as if for the first time. You did not get worse. The record is not on your side of the desk.
This is not a K-12 homework tutor
Say this plainly. HaroCue is not a homework tutor, and it is not aimed at school assignments. The person in mind is a developer, or an adult learning on their own while they ship: someone already moving between ChatGPT, Claude, Cursor, and a codebase. If you want an answer key for classwork, this is the wrong product. If you want to stop re-teaching last week’s conclusion to the next model, keep reading.
Age and the classroom draw a product boundary too. There is no class roster, no homework mode, and no path that submits work for you. Mastery is a decision you make about your own working memory. It is not a grade assigned to a student.
What remembers means
Remembering, here, is not a model’s private impression of you. It is an evidence-backed memory entry. You capture what you understood, in your words, with a source you can open. It becomes mastered only when you confirm it. Until then it can remain an inference, or a question that is still open. HaroCue already separates those: automatic observations are labelled as assistant inferences; confirmed facts stay distinct; you can correct or forget an entry. Ordinary Cue does not treat an unconfirmed inference as an ordinary fact.
The store is local-first on the Mac. Memory lives primarily on the device. That is the same product thesis, with the same limits as the rest of HaroCue: when a cloud model is configured, relevant text can still be sent when you ask or when Cue runs; a local model is available if you want the request to stay on the machine. The record you confirmed is what the next work session can reuse. Inside HaroCue you do not begin from zero. You also do not have to paste the whole story into every new chat you open on that Mac.
Reusable means across work sessions on that machine, not a hidden profile that follows you onto every website. You can later inspect the entry, ask against it, or explicitly have Cursor, Codex, or Claude Code read relevant local context through MCP. Installation does not hand over the whole memory. Scope stays under your control when you make the call.
Different from a generic tutor, and from ChatGPT memory
A generic AI tutor is a conversation with a questioning tone. It can quiz you. It does not owe you a file. After the tab closes, continuity is whatever that vendor kept.
ChatGPT memory is a cloud account feature. The product can retain facts and preferences and bring them into later chats. Inside ChatGPT, that can feel like continuity. It is still memory inside that product. It is not a note you wrote locally, tied to a source, and marked mastered by you. When the hour moves to another tool, that account memory does not become your record unless you carry it yourself.
HaroCue’s difference is those three together: the note stays on the Mac, the source is something you can check, and mastery is a decision you own. It is not a second copy of Study Mode, and it is not a hidden profile the model edits for you. How Study Mode compares with a return to the main question is a separate note.
A workflow you can repeat
Do the questioning where questioning is good. Study Mode is a fine place for that hour. HaroCue is the layer that remembers. It is not another classroom that writes the quiz.
- Study a concept until you can say, in your own sentence, what it does in this project.
- Keep that understanding as a memory, with a source. The source can be the doc you just read, the change you just made, or the chat you already checked.
- Confirm mastery only when you can restate it or apply it on a real change. If you cannot, leave it unconfirmed.
- Before the next drill, look up what you already confirmed, or ask against that memory. Extend the branch you are weak on. Leave the part you already proved alone.
That sequence is the product claim in this article. There is no separate homework mode, no class, and no promise that a model will decide you have learned. The decision is yours.
Start from what you already proved
What HaroCue does today, including sources and boundaries, is on the product overview. The current Mac build is on the download page. The comparison with Study Mode, and how a coder drills down without losing the main thread, are the two notes beside this one.
ChatGPT Study Mode vs local deep learning on Mac
Drill down without losing the thread
Claude for students vs local deep learning