APM-Bench: Benchmarking Cross-session Persistent Memory for Egocentric Streaming Video Assistants
The benchmark tests whether video assistants can remember useful evidence across interrupted, related sessions.
APM-Bench frames egocentric video use as 104 multi-session life trajectories, with 549 sessions and 2,719 candidates. The tasks include objective and open-ended questions, plus proactive responses that depend on past sessions. The paper evaluates general video models and specialized streaming memory systems under different memory protocols. Its finding is that current methods still trade off recall, latency, storage, and proactive usefulness rather than handling all of them well. HF Daily Papers' note
APM-Bench frames egocentric video use as 104 multi-session life trajectories, with 549 sessions and 2,719 candidates. The tasks include objective and open-ended questions, plus proactive responses that depend on past sessions. The paper evaluates general video models and specialized streaming memory systems under different memory protocols. Its finding is that current methods still trade off recall, latency, storage, and proactive usefulness rather than handling all of them well. HF Daily Papers' note
score 5