Agent Memory Challenge Cycle 2 Opens Globally, Inviting More Teams to Benchmark the Future of AI Memory

SAN FRANCISCO, Sept. 28, 2026 (GLOBE NEWSWIRE) -- Agent Memory Leaderboard (AML), an open evaluation platform for long-term Agent Memory, is now accepting registrations for Cycle 2, inviting researchers, developers, open-source teams, commercial AI companies, and individual developers worldwide to participate in a unified evaluation of agent memory systems.

As AI agents move toward longer-running tasks, the ability to retain and use past experience is becoming increasingly important. Yet evaluating Agent Memory remains challenging because systems often rely on different datasets, retrieval methods, prompts, and scoring approaches.

AML provides an open and reproducible framework for comparing Agent Memory systems through a consistent evaluation process. Participants provide standardized Add and Search APIs, while AML handles downstream Answer, Evaluation, result review, and leaderboard publication. Results are reported by track, division, version, and cycle to provide clear evaluation context.

Cycle 1 Builds Momentum for the Next Phase of Agent Memory Evaluation

AML's first cycle drew strong interest from the global AI and developer community, with more than 100 teams participating and dozens of memory systems completing the official evaluation. The AML website recorded more than 300,000 cumulative visits, while its Hugging Face Space reached the top of the Spaces trending list during the cycle.

Building on that momentum, Cycle 2 expands the scope of evaluation to cover a broader range of Agent Memory capabilities and real-world use cases.

Cycle 2 Expands Evaluation Across Three Tracks

Cycle 2 evaluates Agent Memory across Textual Memory, Coding Memory, and Multimodal Memory.

The Textual Memory track covers long conversations and cross-session history, including facts, multi-hop relationships, temporal events, personalization, rule following, governance, and streaming memory. The Coding Memory track evaluates how systems retrieve and reuse historical engineering experience, including implementation context, debugging, failures, testing, and technical decisions. The Multimodal Memory track evaluates systems working with text and image-based memory and evidence.

Cycle 2 includes separate Open-source Methods and Commercial Products divisions. Participation is free and open to researchers, open-source teams, commercial AI teams, and individual developers worldwide.

Eligible open-source teams will compete for more than US$22,000 in total prizes, with US$3,000 awarded to the first-place team in each track.

Evaluation applications and complete materials must be submitted by 23:59 (UTC+8) on October 31, 2026. The current evaluation cycle will close at 23:59 (UTC+8) on November 4, with results expected to be announced in mid-November.

Building an Open Evaluation Layer for Agent Memory

As AI agents take on longer-running tasks, an important question is whether they can effectively use what they have experienced before, not simply access more context.

AML provides an independent evaluation layer for this emerging capability. Participants control their own memory systems and retrieval infrastructure, while AML provides a shared framework for downstream evaluation and comparison.

AML invites researchers, open-source developers, commercial AI teams, and individual developers worldwide to participate in Cycle 2 and contribute to the next phase of Agent Memory evaluation.

Learn more

• Agent Memory Leaderboard
• GitHub

Media Contact：Irene
Email：amlchallenge97@gmail.com