Research map / Video / Video-language understanding
Memento: Toward an All-Day Proactive Assistant for Ultra-Long Streaming Video
Hongxiang Jiang, Zengrui Ge, guo chen, Qixiong Wang, Jile Jiao, Xuetao Feng, Yuan Wang, Yan Wang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add dynamic memory instead of fixed-length memory bank (ablation) | Memento · MementoBench (memory schema ablation) | TimeRecall temporal: 31.2% -> 46.7% | Improved |
| replace fixed memory Len=8 by Len=128 (ablation) | Memento fixed memory · MementoBench (memory schema ablation) | TimeRecall average: 16.9% -> 29.0% | Improved |
| raise dynamic memory threshold from 0.7 to 0.8 (ablation) | Memento dynamic memory · MementoBench (memory schema ablation) | TimeRecall average: 40.4% -> 44.7% | No clear effect |
| lower dynamic memory threshold to 0.6 (ablation) | Memento dynamic memory · MementoBench (memory schema ablation) | TimeRecall average: 40.4% -> 23.8% | Hurt |
| use 1+3x3 frame tokens instead of 1+2x2 (ablation) | Memento · MementoBench (frame token ablation) | TimeRecall: 40.4% -> 68.9% | Improved |
| use 1+4x4 frame tokens instead of 1+3x3 (ablation) | Memento · MementoBench (frame token ablation) | TimeRecall: 68.9% -> 60.9% | Hurt |
| add QMS top-k selection at 50% instead of full memory (ablation) | Memento · MementoBench (QMS top-k ablation) | TimeRecall average: 40.4% -> 56.1% | Improved |
| reduce QMS ratio to 10% (ablation) | Memento · MementoBench (QMS top-k ablation) | TimeRecall average: 56.1% -> 33.6% | Hurt |
| add QMS 50% selection (ablation) | Memento · MementoBench (QMS top-k ablation) | Memory usage: 55.44 GB -> 45.19 GB | Improved |
| add QMS 50% selection (ablation) | Memento · MementoBench (QMS top-k ablation) | Redundancy: 56.2% -> 66.7% | Hurt |
| train VideoLLM-online with Memento-54k (SFT) | VideoLLM-online · MementoBench | TimeRecall average: 8.1% -> 8.9% | No clear effect |
| replace VideoLLM-online* by Memento | VideoLLM-online* · MementoBench | TimeRecall average: 8.9% -> 47.5% | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- VideoLLM-online — VideoLLM-online: Online Video Large Language Model for Streaming Video (CVPR 2024) · better on 5 of 6 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- ReWind: Understanding Long Videos with Instructed Learnable Memory (CVPR 2025 · similarity 0.89)
- LiveStar: Live Streaming Assistant for Real-World Online Video Understanding (NeurIPS 2025 · similarity 0.88)
- Streaming Long Video Understanding with Large Language Models (NeurIPS 2024 · similarity 0.88)
- StreamForest: Efficient Online Video Understanding with Persistent Event Memory (NeurIPS 2025 · similarity 0.88)
- Unleashing Hour-Scale Video Training for Long Video-Language Understanding (NeurIPS 2025 · similarity 0.88)
- TimeChat: A Time-sensitive Multimodal Large Language Model for Long Video Understanding (CVPR 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
