Abstract Reasoning Composition

Concept-level memory beats instance memory: Two formats: Open-Ended (OE) with simple situation → suggestion pairs, and Program-Synthesis (PS) with typed, parameterized routines that support higher-order composition and reuse.

Write = abstract from traces. Read = select with reasoning: OE writes via post-hoc derivations to extract situation/suggestion pairs. PS writes via pseudocode to avoid over-specific details and revises existing concepts. OE selects with a VLM caption and top-k similarity; PS selects with reasoning-based exploration that uses relevance cues and type annotations to decide which concepts to load.

Results on ARC-AGI-1 are strong and scale with retries: With OpenAI o4-mini, ArcMemo-PS lifts the official score from 55.17 → 59.33 on a 100-puzzle subset, a 7.5% relative gain over a no-memory baseline, and remains the only memory design that wins across all tested compute scales. With retries, PS reaches 70.83. See Table 1 on page 8 for the main numbers.

Selection matters for both accuracy and cost: Ablating PS’s reasoning-based selection drops performance and increases tokens. Manual analysis found ArcMemo’s solutions are more attributable to selected concepts than a dynamic cheatsheet baseline that appends all notes.

Continual updates help at scale: Updating memory during evaluation (every few problems) yields additional solves after later passes, supporting test-time self-improvement when verifiable feedback exists.