Led by You, Elevated by AI.
MarginNote's AI isn't a separate mode you switch into — it's embedded across the workflow you already use. Reading, organizing, recalling. You stay in the loop, AI raises the ceiling — but the decisions and the recalling stay with you. Below: where it shows up, and where it deliberately doesn't. Both are design choices.
The map
Grouped by the three learning postures — Read, Write, Recall. Each entry is AI embedded into a specific action you already do, not AI as a parallel "mode" or a separate toolbar.
Auto-extract the structure of a paper or textbook — background / methods / results / discussion — and turn it into a question-driven mind-map skeleton. Every node links back to the exact paragraph in the source. Built on Core ML + Vision + LLM working together.
Multiple modes: smart-dictionary mode (auto-build Title Links for key terms), vocabulary-extraction mode (pull words by CEFR level), question mode (organize by problems instead of by chapter). Two-stage flow — notes first, then organize — breaks past the 100-page limit.
Anywhere in the document or mind map, ask AI in place. Select a passage, hit Ask — the answer is bound to the current passage, no context-switch needed.
Where: document view / mind map view, after selectionStart from any card and let AI respond in "Guide Mode" — instead of dumping full answers, AI asks back, hints, or breaks down steps, leaving the deciding and the connecting to you. Every AI output is something you can Accept or Reject: AI helps expand the mind map, but you stay in charge.
Where: right sidebar (collapsible) / right-click any card → "Talk to AI"When a PDF has no TOC, one click generates a structured, editable TOC — split into Parts / chapters / sections, each with a page range. After generation you can edit it manually and freely switch back to the original.
Where: document sidebar (an "AI generate TOC" button shows up when no TOC exists)
Generate cloze cards directly from your excerpts — fill-in-the-blank plus multiple choice. Right after a chapter, you've got the matching quiz. The gap between "I read it" and "do I actually remember it" closes immediately.
Where: excerpt right-click / card view / review entry
Hand the manual half of card-making to AI — title rewrite (let AI suggest a better card name) and batch polish (rewrite tone and granularity across multiple notes at once).
Where: card right-click / multi-select batch menuLet AI re-arrange your mind map skeleton — based on content relations: re-organize node hierarchy, merge redundancy, flag gaps. When a chapter feels messy, run AI through it once before you commit.
Where: mind map view, right-click "AI restructure"Recognize formulas / tables / handwriting / images / text boxes / vertical classical Chinese / multiple languages — output as Markdown formulas and tables that write directly into notes / mind maps / cards, no copy-paste. Easy to export onward to Obsidian / Notion too.
Max-only: custom prompts let AI auto-organize / translate / summarize / generate flashcards (5 preset templates included), plus a "deep thinking" mode for complex content.
In Recall posture, long-press a blurred excerpt — AI doesn't reveal the answer. Instead it surfaces "same chapter / same branch" related context, helping you associate rather than replacing the act of recall.
This is MarginNote AI's most restrained — and most important — design choice. See "Where AI doesn't show up" below.
Apple Intelligence · iOS 26+
Recall Mode's AI Hint can be generated by Apple Foundation Models right on your device — no network, no AI credits used, your notes never leave your device. Built for libraries, planes, and pre-exam sessions where privacy and offline matter.
"AI Hint model" lets you pick Apple Intelligence (on-device) or cloud LLM. Both paths output the same "one light hint + one concrete prompt" format — you're choosing where the data flows, not what the experience feels like.
Even if you've picked cloud LLM, going offline (planes, subway, libraries) auto-routes to Apple Intelligence — provided your device supports it. The hint loop never breaks just because you have no signal. The reverse doesn't happen: pick Apple AI, your data is never silently sent to the cloud.
Requires iOS 26 + an Apple Intelligence-capable device. Devices / regions that don't qualify automatically stay on the cloud path.
Design choice · deliberately not
"AI isn't in everything we do. It's in the places where it actually helps — and out of the places where it would break active recall."
The Recall-mode AI Hint surfaces "same chapter / same branch" cues only — never the original passage. The moment AI tells you the answer, the testing effect disappears. Engineering restraint, in service of active recall. Cues are assistance; answers are replacement.
Title Link runs on a local dictionary match — no cloud calls, no AI inference, instant response. Shipped in MN 3.6.4 (2019), running for six years. AI here would be slow and pointless — local deterministic matching is the better tool.
The three views (document / mind map / review) share the same object — not "PDF highlight + mind-map node + Anki card, three things, AI keeps them in sync". It's the same object to begin with. Inserting AI here would only add latency and inconsistency.
Pricing
Pro comes with a 500-credit starter — and that starter never expires; it stays in your balance until you spend it. Max adds 500 more every month on top of that starter (while subscribed); the two pools sit side by side. There's one more difference: Pro is locked to the default ("Balanced") model; only Max users can switch among advanced models like Kimi K2, GPT 5, and Gemini 2.5 Flash (Kimi K2 is available in all regions; GPT 5 and Gemini 2.5 Flash are visible only in international regions). Every non-AI capability (card axis, Title Link, Extended Note, Immersive Read/Write, FSRS review) is fully included in Pro.
| Plan | Price | AI credits | AI features |
|---|---|---|---|
| Pro lifetime | $12.99 (iPhone/iPad only) | 500 one-time | Default ("Balanced") model only |
| Max annual | $16.99/year (1-month free trial) | 500 starter + 500/mo | All models (Kimi K2 everywhere; GPT 5 / Gemini 2.5 Flash international regions only) |
| Max lifetime | $51.99 (one-time) | 500 starter + 500/mo (first 3 years) | All models (first 3 years) |
Detailed credit consumption rules, cross-device license sharing, and the "Pro + iOS Pro = Max" equivalence — see the full pricing page →
14-day DMG trial / 1-month Max subscription trial. Pro works as a starting point too — upgrade to Max only if you actually need the AI.
Download free →