What the agent knows about you

The agent explains at the level that you need, not at a fixed beginner or expert level. It finds what you know in your notes. There is no hidden profile of you: all of the evidence is Markdown that you can read, change and delete.

Why the level matters

An explanation that helps a beginner can slow down an expert. Learning research calls this the expertise reversal effect. Thus the agent must know something about you before it explains. It must also not claim to know more than it does. The protocol of the plugin tells the agent to adapt to your knowledge only from evidence that it actually read. If it uses a note as evidence, it must say which note.

The evidence, from weak to strong

Evidence	What it shows	Strength
material in your vault, such as books and clippings	you saw the topic	weak
Your background in the settings	what you think you know	weak
plainer and deeper parts in kept explanations	which depth suited you	weak, but it adds up
notes that you wrote yourself	you may understand parts of the topic	medium
kept recalls and checks: your words, the result and the date	what you really understand	strong

A kept quiz is between these levels: it shows what you recognize, question by question. A recent ✓ counts more than an old ≈.

Who wrote what

The evidence is useful only if the agent can tell your text from its own text. If it read its own notes as your notes, it would see proof of understanding that is not there. So the plugin marks each thing that the agent wrote:

• A note from the agent has generated-by: icap-agent in its properties.
• A kept explanation is a gloss callout.
• A reply in a kept request is in a request callout. Only the Request: lines in it are yours.
• Feedback is in a feedback or practice callout. Your words are outside the callout, directly above it.

The agent counts none of its own text as evidence of what you know. A task note is also not evidence.

Each record has its date and result

A kept recall or check has the date and a mark in its callout title: ✓ holds, ≈ partly, ? look again. The comment above the callout gives the same date and result, and the topic. A kept quiz counts the marks for its questions. Thus each record has its date without the plugin’s data, and the agent can see how old the evidence is.

What the agent gets for a task

For each task, the plugin gives the agent a context from the note:

• the path and the title of the note, and its properties,
• the headings above the focus,
• the paragraphs around the focus, the section and a part of the note around it,
• the links in the note,
• Your background from the settings.

The agent finds more only when it needs it. For example, before it explains a term, it can look at the backlinks of the term and look for your own notes on it. The plugin does not scan the whole vault. It has no vector database and no knowledge graph.

Light diagnosis

The explanation always comes first. The agent does not test you before it explains. If there is no evidence, and the explanation for a beginner is very different from the explanation for an expert, the agent writes for the level of the note. Then it adds one short line for the other level, such as “If you already know X: …”.

The agent judges your level for one concept or one note at a time. There is no global level and no score that follows you.

The ICAP framework in the plugin

The ICAP framework sorts what a learner does into four modes. The deeper modes usually give better learning. The plugin makes each mode easy, but it does not make the wrong thing easy:

Mode	The plugin makes this easy	The plugin does not make this easy
Passive	to understand a difficult term without leaving the paragraph	to replace your reading with the agent’s text
Active	to decide what stays in the note	to save everything automatically
Constructive	to write in your own words without the text	to let the agent rewrite your answer
Interactive	to answer a follow-up question and revise	to drift into an open chat

A mode is what you do, not the name of a control. To select an option is not interactive learning. A conversation is interactive only when you and the agent both reason about the content, reply to each other and revise.

Thus the agent gives a hint or a question for a wrong answer, not the correct answer. A recall, a check or a quiz is always an offer. You can always read and ask without one.