← All research

Continual learning research · the benchmark in one page · November 2026

What Muscle Memory tests that ARC-AGI does not

ARC-AGI measures how well you solve a task you have never seen. It never measures whether solving it changed you. Muscle Memory is built so that it has to.

The benchmark, in one page · November 2026

What Muscle Memory tests that ARC-AGI does not

Muscle Memory, the main benchmark · play it · the two players · no AI run on it yet

ARC-AGI asks: can you solve a new puzzle?

Muscle Memory asks: do you still have yesterday’s trick while you learn today’s?

What we test

  1. Learn by playing

    Nothing is explained. A worker finds the switch.

  2. Remember it

    Next level, the switch is the only way in.

  3. Build on it

    Level 8: two rings, two older moves.

  4. Memory stays on

    Context, notes or weights: your choice.

  5. Gold: 160 leftGold: 160 leftGold: 64 leftGold: 36 leftGold: 180 leftGold: 156 leftGold: 300 leftGold: 152 left
    Weights vs context

    The same AI, played both ways, same board.

  6. Real time

    The enemy knows the move and comes for you.

  7. Gold: 160 leftGold: 160 leftGold: 160 leftGold: 40 leftGold: 188 leftGold: 300 leftGold: 284 leftGold: 16 leftGold: 160 leftGold: 160 leftGold: 160 leftGold: 160 left
    It keeps growing

    More levels until one player falls behind.

What ARC-AGI tests

  • ARC-AGI-1 2019

    Grid puzzles. See a few examples, draw the answer.

  • ARC-AGI-2 2025

    The same, harder. One puzzle at a time.

  • ARC-AGI-3 2026

    Little games, no instructions. Each game scored alone.

Top: ARC tasks sit side by side, each scored on its own, a wall between them. Bottom: Muscle Memory levels form a chain, each one handing the next the move it needs.ARC-AGI: tasks scored alonetask 1task 2task 3task 4…nothing carries over, and nothing checks whether it didMuscle Memory: a chainlevel 1level 2level 3level 4…

Side by side

ARC-1ARC-2ARC-3Muscle Memory
Rules never explained
Learn by acting
Later task needs an earlier one
Memory across tasks is scored
Weights vs context, same model
A live opponent, time counts

yes partly no

ARC-AGI scores the solving. We score what it did to you.

Why the ARC-AGI-3 harnesses should struggle here

They are a frozen model, a long context and a notebook it rereads (agno, executable world models). Our guess, not yet measured:

  1. No line to write down

    Two plates, three seconds. A timing, not a sentence.

  2. The clock runs while it thinks

    Reread the notebook and the enemy is already in.

  3. Gold: 160 leftGold: 160 leftGold: 160 leftGold: 40 leftGold: 188 leftGold: 300 leftGold: 284 leftGold: 16 leftGold: 160 leftGold: 160 leftGold: 160 leftGold: 160 left
    The context overflows

    A whole board twice a second. Summaries drop the exact bit.

  4. It restarts every level

    From notes, not from skill. The bell lifts the gates for eight seconds.

Build it yourself

Build your own copy of this testTwelve rules; click one to read it. The button copies the full prompt as plain text, ready to paste into your agent.

Click a rule to read it in full.

If they keep up anyway, our bet is lost. Either way it is a number ARC-AGI never takes. Run one.

More research