Multimodal Agent MemorySubmit a paper

HD-EPIC Long-Video Reasoning

Semantic and Visual Evidence for Efficient Long-Video Reasoning: A Solution for the HD-EPIC VQA Challenge

Yinsong Xu, Wei Jing, Liuxin Zhang, Wanjun Lv, Hui Li · 2026-05-28

Read the paperPDF

Management
T1 Static Memory
Memory is built once and reused without change during deployment.
Primary modality
Video & streaming
Also: text, Image
Representation
Source- and modality-specific

How experience becomes memory

Each path is one memory component, traced from what is stored to how it is found again.

  1. selected semantic and visual evidence relevant to each question

    1. Represented asRetained perceptual source evidenceSource- and modality-specific
    2. Stored inFlat sequential records
    3. Retrieved byPlanner / tool-assisted retrieval
    4. Source evidenceSource directly available
    Evidence representations (text-form semantic cues + visual frame cues) selected per question. question-conditioned evidence selection balances semantic and visual signals.Full paper: 3.1. Implementation Details · PDF p. 3 · Figure 1 / method overview

Related systems