Multimodal Agent MemorySubmit a paper

MangaFlow

MangaFlow: An End-to-End Agentic Framework for Controllable Story to Manga Generation

Muyao Wang, Zeke Xie, Yanhao Chen, Lixin Xiu, Hideki Nakayama · 2026-05-27

Read the paperPDF

Management
T2 Dynamic Memory
Memory changes during deployment through a stable management policy.
Primary modality
Image
Also: text
Representation
Hybrid
The cached reference images and the textual section description / structured profiles live in one section record and are jointly supplied to the panel renderer, which conditions on the images while being prompted with the text profiles.

How experience becomes memory

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

  1. story-section memory: cached visual reference images for shared scenes, characters and key objects (R_k^scene, R_k^char, R_k^obj)

    1. Represented asRetained perceptual source evidenceSource- and modality-specific
    2. Stored inFlat sequential records
    3. Retrieved byMetadata / spatiotemporal retrieval
    4. Source evidenceSource directly available
    Each story section s_k owns a memory record keyed by section identity. The record holds generated or user-supplied reference IMAGES for the section's shared scenes, characters and key objects. Algorithm 1 loads the cached memory M_k if available and otherwise generates and stores the references; every panel assigned to section k then reads the same stored images and conditions rendering on them to preserve scene/character/object consistency. Access is by section (and entity) key, not by embedding similarity. The stored images are the artifacts themselves, so the perceptual source is directly available downstream.Full paper: Sec. 3.4 Story Section Memory; Sec. 3.6 reference-conditioned rendering · arXiv HTML, method section · Equation 6 (M_k = (d_k, R_k^scene, R_k^char, R_k^obj, phi_k)); Algorithm 1 lines 4-10

  2. story-section memory: textual section description d_k and structured character / key-object profiles phi_k

    1. Represented asPerception-to-text records & summariesCross-represented explicit
    2. Stored inFlat sequential records
    3. Retrieved byMetadata / spatiotemporal retrieval
    4. Source evidenceSource not retained
    Alongside the reference images the same record persists a textual section description d_k and structured profiles phi_k describing characters and key objects. These text records are written once per section and are re-read as prompt conditioning for every panel in that section, so they are an explicit cross-represented text memory distinct from the stored images.Full paper: Sec. 3.4 Story Section Memory · arXiv HTML, method section · Equation 6; Algorithm 1 lines 4-10

How memory changes (T2 operations)

  • Update · Rule-based update

    Memory is populated lazily inside the generation loop rather than fully built up front: for each section the framework first tries to load the cached M_k, and only if it is unavailable does it generate the references and write the new record. The store therefore grows during generation under a fixed deterministic write rule, and later panels read entries written by earlier iterations. No revision, merge, or eviction of existing entries is described.

Related systems