WorldMemArena

WorldMemArena WorldMemArena is a unified evaluation suite for long-horizon, multimodal memory in agent and lifelong-assistant settings. It bundles 461 long-running interaction samples across 27 sub-domains, together with all referenced screenshots, trajectory frames, and embedded media required to reproduce evaluation. Each sample contains: a multi-session dialogue history (avg. 18 sessions / sample, up to 30 turns per session), a structured memory point stream describing what an… See the full description on the dataset page:

Source: Hugging Face Hub (WorldMemArena/WorldMemArena). Metadata imported from the dataset’s Hub tags.