MemDreamer: Decoupling Perception and Reasoning for Long Video Understanding via Hierarchical Graph Memory and Agentic Retrieval Mechanism
Summary
MemDreamer decouples perception and reasoning for long video understanding using hierarchical graph memory and agentic retrieval, achieving state-of-the-art performance with reduced computational overhead.
View Cached Full Text
Cached at: 06/10/26, 09:44 AM
Paper page - MemDreamer: Decoupling Perception and Reasoning for Long Video Understanding via Hierarchical Graph Memory and Agentic Retrieval Mechanism
Source: https://huggingface.co/papers/2606.07512
Abstract
MemDreamer addresses long-video understanding challenges by decoupling perception and reasoning through hierarchical graph memory and agentic exploration, achieving state-of-the-art performance with reduced computational overhead.
CurrentVision-Language Modelsstruggle with hours-long videos because processing full-length visual sequences induces prohibitivetoken explosionandattention dilution. To overcome this, we introduce MemDreamer to decouple perception and reasoning, shifting long-video understanding into anagentic explorationprocess. As a plug-and-play framework, it incrementally streams videos to construct aHierarchical Graph Memory, a top-down three-tier architecture for semantic abstraction, anchored by a foundational graph capturing spatiotemporal andcausal relations. During inference, the reasoning model employs agentictool-augmented retrieval, navigating hierarchies, searching nodes, and traversing logical edges via anObservation-Reason-Action loop. Experiments show MemDreamer achieves SOTA results across four mainstream benchmarks, narrowing the gap with human experts to only 3.7 points. It constrains the reasoning context window to merely 2% of full-context ingestion while delivering a 12.5 point absolute accuracy gain. Furthermore, statistical analysis uncovers a strong positive linear correlation between an VLM’s performance on logic reasoning and long-video understanding benchmarks, establishing agentic capability scaling as a new paradigm formultimodal comprehension.
View arXiv pageView PDFProject pageAdd to collection
Models citing this paper0
No model linking this paper
Cite arxiv.org/abs/2606.07512 in a model README.md to link it from this page.
Datasets citing this paper0
No dataset linking this paper
Cite arxiv.org/abs/2606.07512 in a dataset README.md to link it from this page.
Spaces citing this paper0
No Space linking this paper
Cite arxiv.org/abs/2606.07512 in a Space README.md to link it from this page.
Collections including this paper0
No Collection including this paper
Add this paper to acollectionto link it from this page.
Similar Articles
Reasoning with Memory: A Temporal Granularity-Adaptive Framework for Training-Free Long Video Understanding
ReMem introduces a dual-level memory-augmented keyframe selection framework for training-free long video understanding, achieving state-of-the-art zero-shot performance on multiple benchmarks.
MemLife: Curating and Reasoning over Long-Term Egocentric Video Memories
MemLife introduces a multimodal memory system for long-term egocentric videos that builds entity-grounded text episodes and retrieves them with a time-indexed agentic reader, improving over training-free baselines by 4.6-12.0% on long-horizon benchmarks. A reinforcement learning framework called MemOpt further optimizes the memory writer for faithfulness and retrievability, yielding consistent 2.7-5.0% gains.
Memory is Reconstructed, Not Retrieved: Graph Memory for LLM Agents
MRAgent introduces a novel graph-based memory framework that dynamically reconstructs memory during reasoning, achieving up to 23% improvement on long-horizon benchmarks while reducing computational costs.
GraMRAG: Orchestrating Multi-Agent Multi-Step Reasoning via Graph Memory with Reinforcement Learning
GraMRAG is a new multi-agent RAG framework that uses graph memory and reinforcement learning to improve reasoning depth and memory structure in complex multimodal tasks, achieving state-of-the-art performance.
RRM: Experience-Driven Reflective Retrieval Memory for Long-Horizon Multimodal Reasoning
This paper introduces Reflective Retrieval Memory (RRM), a memory framework that distills procedural retrieval experience from historical task trajectories to improve evidence retrieval for long-horizon multimodal reasoning. RRM matches or exceeds prior state-of-the-art on M3-Bench-Robot, M3-Bench-Web, and Video-MME-Long benchmarks.