Agent Memory Challenge 2026 Cycle 2 Opens September 20

Summary

The Agent Memory Leaderboard (AML) will open Cycle 2 of its Agent Memory Challenge on September 20, 2026, at 00:00 UTC+8. The program evaluates whether long-running AI agents can retrieve current, useful evidence from past experience, rather than merely store more history or repeat outdated facts. Cycle 2 covers Textual Memory, Coding Memory, and Multimodal Memory. The textual track tests cross-session recall, multi-hop relations, temporal events, personalization, rule following, governance, and a new Streaming Memory setting in which information arrives over time. The coding track evaluates whether repository history, debugging trails, failure cases, module relationships, tests, and prior decisions help agents solve 150 software-engineering tasks under both relevant-history and noisy-history conditions, for 300 task-condition units. The multimodal track tests memory built from text, images, captions, or combinations of these, including preservation of ordered text-and-image content. Participants provide stable Add and Search APIs. Add must persist content before returning success, while Search must return traceable evidence rather than a final answer; AML runs the downstream Answer and Eval steps using a common protocol. The evaluation also requires strict user isolation, immediate searchability of newly written memory, idempotent request IDs, fixed submitted versions, and stable operating conditions. Cycle 2 is open worldwide to academic, open-source, commercial, independent, and cross-organization teams, with separate Open-source Methods and Commercial Products divisions. Participation is free, but teams cover their own infrastructure. Eligible open-source teams will compete for a total RMB 150,000 prize pool across the three tracks. Registration and evaluation run from September 20 through October 31, materials are due October 31, evaluation closes November 4, and official results are planned for mid-November, subject to the latest organizer rules.