APM-Bench: Benchmarking Cross-session Persistent Memory for Egocentric Streaming Video Assistants
APM-Bench evaluates persistent memory for streaming video assistants across 104 multi-session trajectories.
Researchers introduce APM-Bench to test persistent memory for egocentric streaming video assistants across interrupted, multi-session life trajectories rather than single clips. The benchmark contains 549 annotated sessions, 104 trajectories, and 2,719 objective and open-ended candidates. Models must store memory selectively, inject it at the right time, answer about earlier sessions, and recognize when evidence is unavailable. Tests of general video models and specialized streaming memory systems show they still cannot combine reliable long-term recall, low latency, and efficient storage.
- Benchmark has 549 sessions, 104 trajectories, and 2,719 candidates.
- Sessions within a trajectory cover related but interrupted activities.
- Models must recall past sessions and admit missing evidence.
- Existing methods still miss the recall, latency, and storage balance.
From the original article
1 min read at huggingface.co
To serve as real-world personal assistants, streaming video models need persistent memory that retains past experiences for later use. Yet existing streaming benchmarks and methods often focus on individual continuous videos or short clips, overlooking that real-world interactions are often intermittent and require memory to persist across interruptions.
By Jianguo Huang, Jinming Liu, Qiyao Wang · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.