It writes real blocks
Text, live LaTeX, a plotted graph, a table, a runnable code cell — all native, all editable, each measured against the canvas so nothing overlaps.
Meet Lemma AI
AI that reads
AI that reads
your notebook
It works inside the note you are already in — the equation you selected, the lecture you captured this morning — and writes its answer back onto the canvas as blocks you can edit.
72.9%Pass rate, against 37.5% raw
8Review stages before you see it
5Block types it writes
What it does
Everything ends up on your canvas
It already has the context
Select boxes and they attach to your next message. Ask “why does this step work?” and this means the equation under your cursor.
Photograph what you missed
A line from the textbook, a step left on the board. Vision transcribes the maths off the page and the explanation lands in the note you are already in.
It finds the first wrong step
A line number, the rule you skipped, and why it matters — not a rewritten solution to copy. It is tuned to teach the method.
It sits through the lecture
Board photos, a posted PDF, or a recording of the hour. It names each stage as it runs — read, write, check, save — and Cancel actually stops it.
It follows you between devices
A note written on your phone is re-measured and re-laid out the first time it opens on a desktop canvas. No signal in the hall? It uploads when you reconnect.
Try it
It quizzes you from your own notes
One /quiz turns the lecture you just wrote into practice questions with the answers hidden, so revision is recall rather than rereading.
The characteristic polynomial. Its roots are the eigenvalues of A.
Exactly when it has n independent eigenvectors — enough to form a basis.
Only if λ = 2 has two independent eigenvectors. The repeated root is the case worth checking.
Two dials
Effort decides how hard it thinks. Intent decides what it builds.
Independent on purpose — the same question at Max effort gives a study schedule under Plan and a mind map under Brainstorm.
Effort
- FastReasoning off — a definition, a unit conversion, a sanity check.
- AdaptiveDefault. Balances speed against quality per task.
- MaxDeep reasoning for the derivation you have been stuck on.
Intent
- NotesStructured notes you can study from.
- BrainstormA mind map that spreads across the canvas.
- PlanA dated, checkable study schedule.
- QuizPractice questions with hidden answers.
- RecapThe short version — key points only.
Or type a slash
- /notesFull canvas notes on a topic.
- /solveOne problem, worked step by step.
- /explainThe attached boxes, in plain terms.
- /checkHunt the working for mistakes.
- /quiz · /summaryPractice questions, or a review sheet.
A day with it
From the lecture hall to the exam
09:15
In the lecture
Photograph the board, or start a recording and keep listening. Capture is the only job.
09:58
Walking out
Already transcribed and written up — title, definitions, key formula, one worked example.
14:00
At the desk
The note re-lays itself out for the bigger screen. Select the step you didn’t follow and ask.
21:30
The night before
/quiz the week, /summary for a review sheet, a dated plan for what is left.
Review mode
It marks its own work before you see it
The failure that matters here is a confident wrong number in a box you will revise from. So every reply is re-solved from the original problem by a model that never saw the draft.
**Step 1** — find the cash flows
IRR = 9.2018Off by two orders of magnitude — and it lints cleanly, because a wrong number is still a number.
$$\text{NPV} = 0$$ at r = 0.0918Contradicts the box above it. Nothing in a plain pass compares the two.
Sheet opens a fourth column, off the edge of the canvas.
Step 1 — find the cash flowsLocal repair, free: stray markdown stripped, rebuilt as a proper text box.
IRR = 9.18%A local calculator proved the two displayed values could not both hold.
NPV = 0 at r = 0.0918The independent solve agreed at 0.0918, so the draft was patched to match.
Repacked to three columns — the sheet fits the viewport.
The pipeline
Eight stages, each with a trigger and a rollback. No stage may make the sheet worse, and the free ones run whether the toggle is on or off.
- Local repairUnformatted prose, missing or duplicated coordinates — fixed in the browser.free
- Sentence rowsA lead-in written after the formula it introduces, reordered so the row reads.free
- StructureOnly the commands still failing the lint are sent back to be rewritten.1 call
- ContradictionFires when a local calculator proves two values cannot both be true.1 call
- Independent solveSolved again from scratch by a model that is never shown the draft.1 call
- ConfirmA second, independent solve must agree before anything may be edited.1 call
- PatchOnly now is the box rewritten — and only the boxes that were wrong.1 call
- RepackStops an avoidable fourth column running off the canvas. Pure geometry.free
The evidence
What the pipeline is worth, measured
Benchmark · 48 graded trials
Base model vs. review mode
The same model, the same 48 trials. The only difference is whether the review pipeline runs first.
Base model
With review mode
Overall pass▲ +35.4Every check on a trial passing at once
Maths correct▲ +8.3The numbers and the working are right
Layout usable▲ +37.5The sheet fits the canvas and reads in order
Agentic pass▲ +10.4It did the thing that was actually asked
Formatting clean▲ +37.5No stray markdown, no broken LaTeX in a box
One grader, one pass, so the columns are directly comparable. “Base model” is the same model with review switched off.
Try it on a hard derivation
Review is on by default · the free stages always run
Honest by design
Tuned to teach the method, not hand over answers for you to copy.
Your notes stay yours
Private by default. Your work is never used to train shared models.
Cites its sources
When it pulls from your notebook, it tells you which note it used.
Questions
The ones people actually ask
Can I turn review mode off?
Yes — the toggle in the composer. It silences the model stages only. Local repair, row alignment and repacking always run, because they cost nothing and they are the difference between a tidy sheet and one that runs off the side of the canvas.
Does review make it slower?
Only when it fires. Most replies clear the local lint and never reach a model stage. When a quantitative draft does trigger a solve, expect a few extra seconds — and the whole pipeline runs under one overall budget.
Will it just do my homework?
It will solve a problem if you ask, and it will show every step while doing it. What it is tuned against is the bare answer with no working — the version you cannot reproduce in an exam.
What happens to my notes?
They stay yours. Notes are private by default and are not used to train shared models. Captures are processed to build the note you asked for and nothing else.
Does it work on my phone?
Yes — the same pipeline runs on both. The desktop places boxes on an infinite canvas; the phone re-joins them into a scrolling document with inline maths. The stored note is identical either way.
Study with an AI that has the context
Lemma AI is included free in the beta. Bring a derivation you have been stuck on and let it explain.
Get early access