AML Tests Agent Memory Under Noisy Context

AML focuses on the metric that matters: whether retrieved history helps an agent ship correct code, not merely whether the retrieval system finds similar text.

• –300 scored attempts expose how much noisy context degrades coding-agent performance
• –Tasks span real repositories, including both feature work and bug fixes
• –The benchmark separates memory retrieval from the downstream coding agent for cleaner comparisons
• –Results should push memory systems toward filtering, freshness, provenance, and actionability
• –Teams building agent harnesses can use this as a stronger test than Recall@K alone