Persistent learner notes: remember what the learner struggles with across sessions

[hieudepzai14122007-ship-it]

Related: #206 (Learner Profiles/Facts).

I built a working version of part of this: remembering what the learner gets wrong across sessions. Before opening a PR I'd like to check that the approach fits, since the README says on purpose that the graph stores ideas, "not ratings of your ability".

Problem. The graph records what was taught, not how the learner is doing. A new session can't see that I keep making the same mistake (for example, forgetting the factor n in the power rule), so it doesn't target it.

What it does — branch · diff

• One app-wide Markdown note with three sections, Struggling with, Improving and Solid now, with a date on each item.
• The prober, teacher and practice agents get an update_learner_notes tool and read the current note in their system prompt. The prompt block is only added when the agent's tool list includes the tool, the same rule other conditional tools follow.
• Each update appears as a collapsed tool-call row in the transcript.
• Settings → Agent → Learner notes lets you review, edit or clear it.
• It's the model's written judgment, not a score, and nothing is added to the graph.

Verification. I added tests for the tool and the prompt composition, and updated the tests that pin each agent's tool list. The full suite shows no new failures. The 19 failures I see are Windows-specific and fail on main too. Live check with Claude Code (Opus): I answered three probe questions with the "just lower the exponent" misconception, and the prober wrote an accurate note before planning.

Open questions

• Should this be core, or wait for the plugin infrastructure?
• One global note or per-topic notes? Should it also hold profile facts like learning preferences, as Learner Profiles/Facts #206 suggests?
• Should it be off by default, or a toggle under Tools?

I'm happy to adjust it, or open a PR if the direction works.

[AurelianVIII]

Thanks for building this. I'd like to extend it to cover the other half of #206: how the learner learns (pace, what makes an explanation click), not only what they get wrong. That changes more slowly than struggles, so I think it needs different update rules, such as end-of-session updates with evidence for each line and a size budget. I run a similar extract-and-merge loop for my own tutor prompts, so I've already hit those problems.

If the direction works for the maintainers, I'd like to build that section on top of your branch. I can start in about four weeks.