UCLA Introduces LongMemEval-V2: Benchmarking Agent Environmental Memory

dair_ai · x · 2026-08-27

Most agent memory benchmarks test user history recall, but production agents need to internalize environmental characteristics like interface quirks and state dynamics. A new UCLA paper introduces LongMemEval-V2, a benchmark with 451 manually curated questions covering five web agent memory abilities: static state recall, dynamic state tracking, workflow knowledge, environment gotchas, and premise awareness. History trajectories scale up to 500 traces and 115M tokens per question.

The proposed AgentRunbook-C method stores trajectories as files, allowing a coding agent to gather evidence in a sandbox. It achieves an average accuracy of 72.5%, significantly outperforming the strongest RAG baseline (48.5%) and an off-the-shelf coding agent (69.3%).

Original post →

More from coding & agent

coding & agent channel →