Bridging Modalities, Spanning Time: Structured Memory for Ultra-Long Agentic Video Reasoning

Researchers introduce MAGIC-Video, a training-free framework that enables multimodal AI systems to process and reason about ultra-long videos spanning days or weeks by combining a structured memory graph with narrative chains. The system outperforms existing baselines on multiple benchmarks, addressing a critical limitation where current LLMs can only handle tens of minutes of video despite having million-token context windows.

MAGIC-Video tackles a fundamental constraint in video understanding: the inability of current multimodal language models to retain and reason over extended temporal sequences. While state-of-the-art LLMs feature enormous context windows, their frame-sampling strategies discard most visual information before inference, making analysis of surveillance footage, egocentric recordings, or livestreams impractical. This research introduces a dual-architecture approach combining a typed memory graph with narrative chains, enabling both cross-modal and temporal retrieval without requiring model retraining.

The technical innovation addresses fragmentation across modalities and time. Traditional approaches retrieve visual content separately from semantic or episodic information, failing to capture long-range patterns like recurring activities or entity biographies spanning weeks. MAGIC-Video's six typed edges unify these modalities while its narrative chain distills key events and entity arcs, creating a coherent knowledge structure. The agentic inference loop dynamically interleaves graph queries with fact injection, effectively spanning both dimensional axes of ultra-long video.

The practical implications extend across security, healthcare, and autonomous systems. Surveillance applications requiring pattern detection across days now become tractable. Egocentric AI assistants can maintain coherent understanding of users' lifespans. The 10.1-point improvement over prior agentic baselines on EgoLifeQA demonstrates substantial progress. Since the framework operates training-free, adoption barriers remain low—institutions can integrate it with existing LLM infrastructure immediately.

Future development hinges on memory scaling efficiency and handling even longer temporal horizons. Real-world deployments will test whether the structured approach generalizes beyond research benchmarks to messy, unlabeled video streams.

• →MAGIC-Video enables multimodal LLMs to reason over ultra-long videos spanning days/weeks through structured memory graphs and narrative chains.
• →The training-free approach achieves 10.1-point improvements on benchmarks without requiring model retraining, lowering adoption barriers.
• →Cross-modal retrieval and temporal narrative chains address fragmentation in current video understanding systems.
• →Applications span surveillance, egocentric AI, security analytics, and any domain requiring pattern detection across extended timescales.
• →Open-source code availability accelerates industry adoption and potential integration with existing LLM infrastructure.