The Agent Memory Leaderboard (AML) has launched Cycle 2 of the Agent Memory Challenge to evaluate long-term memory capabilities in AI agents, focusing on the ability to retrieve relevant, current evidence rather than just historical similarity. Unlike standard benchmarks, AML separates the memory layer from the reasoning layer; participants provide "Add" and "Search" interfaces, while the AML orchestrates the downstream "Answer" and "Eval" processes to ensure standardized, reproducible comparisons. The evaluation is divided into three distinct tracks: Textual Memory, which tests temporal reasoning and multi-hop relations; Coding Memory, featuring 150 software engineering tasks (300 task-condition units) across relevant and noisy history settings; and Multimodal Memory, which assesses the integration of text and visual context. Cycle 2 includes separate divisions for Open-source Methods and Commercial Products. Following a successful Cycle 1 involving 67 evaluated systems, Cycle 2 offers a prize pool of approximately USD 22,400 for open-source contributors. The evaluation window runs from late September through early November 2026, requiring stable APIs that adhere to strict principles of user isolation, idempotency, and successful write-to-search persistence.