Multimodal Agent MemorySubmit a paper

MemSlides

MemSlides: A Hierarchical Memory Driven Agent Framework for Personalized Slide Generation with Multi-turn Local Revision

Ye Jin, Yangyang Xu, Jun Zhu, Yibo Yang · 2026-06-15

Read the paperPDF

Management
T2 Dynamic Memory
Memory changes during deployment through a stable management policy.
Primary modality
Image
Also: text
Representation
Hybrid
Structured profile and slide-edit state is joined with compact tool-experience records: Plan-Act-Guard reads both paths when planning and executing personalized slide generation and local visual revisions.

How experience becomes memory

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

  1. persistent user profile plus evolving session and slide-edit working state

    1. Represented asSpatial, temporal & cross-modal structured memoryCross-represented explicit
    2. Stored inHierarchical multi-tier store
    3. Retrieved byMetadata / spatiotemporal retrieval, Structure traversal
    4. Source evidenceSource not retained
    The framework maintains a persistent user profile containing visual-style, layout, and template preferences together with an evolving session state. Working memory records temporary preferences, carryover constraints, edit targets, coverage, and snapshot hashes; the planner and actor read this state and the guard updates it after each local slide revision.Full paper: Sec. 3.1 Stateful Multi-turn Slide Authoring; Sec. 3.2 Plan-Act-Guard · arXiv HTML, method section · Figure 2
    Profile memory persists across jobs while active working memory changes over revision rounds through append and supersede operations. At job completion, stable signals are consolidated into the long-term profile for use in later slide-generation jobs.Full paper: Sec. 3.3 Hierarchical Memory · arXiv HTML, method section · Figure 3; Equations 3-6

  2. compact lessons, tool-error summaries, and reusable operation-context fragments

    1. Represented asPerception-to-text records & summariesCross-represented explicit
    2. Stored inHierarchical multi-tier store
    3. Retrieved byPlanner / tool-assisted retrieval
    4. Source evidenceSource not retained
    Raw reasoning, tool-call, and observation chains are segmented into compact reusable fragments indexed by operation context. Relevant fragments are retrieved before similar future tool calls, while transferable round-level lessons and error patterns are consolidated into long-term experience memory.Full paper: Sec. 3.4 Experience Memory · arXiv HTML, method section · Figure 4; Equation 7

How memory changes (T2 operations)

  • Update · Rule-based update

    After each user feedback round, fixed append and supersede rules update the active preference, carryover, target, coverage, and snapshot state that will govern the next local revision.

  • Consolidation · Experience abstraction

    The system transforms raw reasoning-tool-observation traces and round outcomes into reusable fragments, lessons, tool-error summaries, and operation patterns for later retrieval.

  • Reorganization · Memory-level migration (short- to long-term)

    At job completion, stable preference signals and transferable tool experience are consolidated from active working memory into persistent profile and long-term experience tiers for future slide-generation jobs.

Related systems