Start familiar — then level up

The gym
for your mind.

Bring the flashcards you already use — import your Anki deck and practise the way you always have. Or write your own — about whatever you're trying to learn, or how you want to handle it: the exam, the language, the pitch, the conversation you can't afford to fumble. MemGym quietly learns what you actually know, and shows you when you're ready to level up.

Already have access? Sign in →

Closed early access.

2 of 10study methods earned top marks for durable learning — practice testing and spacing (Dunlosky et al., 2013)

0spaced-repetition tools track whether you knew you'd recall — only whether you did

~20 repsand your own calibration curve emerges — from predictions you already made

Anki knows what you remember. MemGym knows where you're wrong about that.

Anki (and every other spaced-rep tool)

Tracks recall: did you get the card right?
Schedules reviews based on past accuracy
Tells you what to study next
Assumes you know how confident you are
No signal on metacognitive accuracy
Recall cards only — no way to rehearse a behaviour

MemGym

Tracks recall plus your stated confidence
Computes your calibration curve in real time
Surfaces domains where you feel sure but aren't
Forces an honest prior before every answer
Measures the gap — and closes it deliberately
Scores behaviour, judgement and reflection too — not only recall

Overconfidence isn't a character flaw. It's an untracked metric. Once it's tracked, it moves.

The calibration loop

Four steps. Repeatable. Measurable.

01
Predict
Before the answer shows, commit to how sure you are — low, medium, or high (1 · 3 · 5). One tap, every card. Never skipped.
02
Test
Reveal the answer and score how it actually went. MemGym records your prediction and your outcome separately — that's the whole trick.
03
Gap
The gap between the two is the signal. Say “high” and miss it, and the gap shows — tracked across every rep, per context.
04
Reflect
Your calibration view shows the trend, whether you lean over- or under-confident, and which context to practise next. The gap is the number that moves.

Your agent can practise with you

MemGym speaks the Model Context Protocol. Connect Claude Code, or Claude with a custom connector, and it works on your actual deck — reading your cards, turning what you are studying into new ones, and running real reps with you. One approval in the browser; no API key to generate or paste.

And it cannot flatter you. Recording a rep takes both your prediction and the outcome, so an agent has no way to log “got it right” without the number only you can give it. A calibration figure means the same thing however you practised.

How to connect an agent →

Built for anyone who has to be right about themselves

Anyone with an exam coming

Finals, the Bar, CFA, medical boards, the theory test. You're already drilling. Add the layer that shows whether your drilling is building real confidence or false confidence.

Language learners

A word you can recall in a quiet room is not a word you can recall in a conversation. That difference is exactly what MemGym measures.

Changing how you act

Not just what you know. Assertion cards are specific, observable behaviours you rate yourself against — predict how true it is today, rehearse it, then score what actually happened. Scenario cards do the same for a situation you are about to walk into: write your move first, then compare it with a strong response. The graph shows whether your read on your own behaviour is getting sharper.

High-stakes work

Physicians, engineers, founders, operators. Overconfidence has a direct downside, and knowing where your knowledge is soft — before a decision costs you — is a real edge.

Deep domain knowledge

Researchers, analysts, or simply someone who has read a great deal about one thing. MemGym tells you which parts of it you actually hold correctly.

What the research actually says

Calibration training has a real literature. The clearest findings:

The feeling of “I know this” is a poor guide to what you'll recall. Confidence during study rides on how fluent material feels right now, not on whether it'll come back later (Koriat & Bjork, 2005). Predicting, then testing, is what exposes the difference.
Calibration is domain-specific — so we only claim it for what you practise. Feedback sharpens calibration inside the trained material, and transfers only weakly beyond it (Moore & Healy, 2008; Good Judgment Project). MemGym maps the domain you're actually drilling — not your judgment in general.
Brief, feedback-driven practice measurably improves calibration. The forecasting literature — the Good Judgment Project (Mellers et al., 2015) — is the clearest signal: short calibration training moved accuracy and held up over time. We borrow the principle, not the domain: it was event forecasting, not memory.

The gap between where you land and the diagonal is the only number that matters.

We will not claim MemGym is clinically validated. It isn't yet. We will show you your numbers and let you decide if they're useful. That's the deal.

Bring your flashcards or write your own cards.

In early access. Import an Anki deck and practise the way you already do, or write your own cards about whatever you're trying to learn — and watch your calibration curve emerge.