AML Benchmark for Agent Memory Released: MemoraX Tops, NetEase Third
机器之心 · wechat · 2026-08-14
The inaugural Agent Memory Leaderboard (AML), co-launched by nearly 30 institutions including Oxford and Tsinghua, has been released to establish a unified evaluation standard for Agent long-term memory systems.
Key Leaderboard Insights
- Industrial Track: MemoraX dominates with a score of 58.0, ranking first across all 7 text memory dimensions. MemOS takes second place, while NetEase's NTES-MEMORY-SMART emerges as a dark horse in third, showing exceptional personalization capabilities.
- Open-Source Track: InvMem, ReFind, and ActiveMemoryIndex take the top three spots, representing distinct technical approaches like hybrid retrieval, iterative search, and fine-grained query rewriting.
Evaluation Mechanism
AML strictly isolates memory interfaces from generation, integrates over 10 benchmark datasets, and employs a multi-agent judging system to ensure fairness. This release marks the transition of the Agent memory track towards standardized evaluation.
More from coding & agent
- DashClaw: Open-Source Approval Layer Intercepts Risky AI Agent Actions in Real Time — SIGH_I_CALL · 2026-08-14
- Axon: Open-Source Tool Turns Codebases into Knowledge Graphs for AI Agents — tom_doerr · 2026-08-14
- Open-Sourcing CoRe: Whole-Body Motion Retargeting Pipeline for 11 Humanoid Robots — rsasaki0109 · 2026-08-14
- Letting Agents Talk Directly: Auto-Fixing Bugs on Both Ends — dinasaur_404 · 2026-08-14
- Open-Sourcing CoRe: Whole-Body Motion Retargeting Pipeline for 11 Humanoid Robots — rsasaki0109 · 2026-08-14
- System 0xF0: Zero-Dependency Agent Coordination Framework with Native /llms.txt Support — KidneeBean · 2026-08-14