APM-Bench

APM-Bench: Benchmarking cross-Session Persistent Memory for Real-World Egocentric Streaming Video Assistants

Paper · Project page · Code

APM-Bench groups egocentric video into activity-related trajectories that span separate interactions. Its tasks test whether an assistant can use earlier experience while answering questions, perceiving the current scene, and deciding when to respond.

Dataset statistics

Video source	Trajectories	Sessions	Candidates
EgoLife	76	472	1,896
HD-EPIC	28	77	823
Total	104	549	2,719

Capability	Candidates
Cross-session understanding	887
Real-time perception	716
Adaptive response	1,116

The benchmark covers 12 tasks and 5,371 evaluation probes. The counts above describe the complete benchmark; this repository currently contains its 549 session videos.

Release plan

• Initialize the dataset repository and code repository.
• Release the dataset and code before November 2026. We are preparing the annotations and evaluation code now.