APM-Bench: Benchmarking Cross-session Persistent Memory for Egocentric Streaming Video Assistants

Real-world video assistants need persistent memory to retain past experiences across intermittent interactions, a requirement overlooked by continuous streaming benchmarks. The authors introduce APM-Bench, reformulating real-world streaming into multi-session life trajectories comprising 549 sessions, 104 trajectories, and 2,719 candidates with fine-grained annotations. Evaluating general video models and specialized memory systems on APM-Bench reveals a clear utility-latency-storage trade-off, demonstrating that existing methods struggle to achieve reliable long-term recall, low overhead, and effective proactive assistance simultaneously.

APM-Bench contains 549 sessions, 104 trajectories, and 2,719 candidates spanning objective and open-ended questions.

Evaluation reveals a clear utility-latency-storage trade-off where existing methods struggle with reliable long-term recall and low overhead.