Multimodal Agent MemorySubmit a paper

SymbOmni

SymbOmni: Evolving Agentic Omni Models via Symbolic Concept Learning

Jinxiu Liu, Jianru Li, Tanqing Kuang, Xuanming Liu, Kangfu Mei, Yandong Wen, et al. · 2026-07-13

Read the paperPDF

Management
T3 Self-Evolving Memory
Feedback during deployment changes the memory-management policy itself.
Primary modality
Image
Also: video, text
Representation
Cross-represented explicit

How experience becomes memory

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

  1. persistent symbolic concept repository for visual creation

    1. Represented asSpatial, temporal & cross-modal structured memoryCross-represented explicit
    2. Stored inVector-indexed collection, Graph / relational store
    3. Retrieved bySimilarity search, Structure traversal, Planner / tool-assisted retrieval
    4. Source evidenceSource not retained
    The repository stores symbolic visual concepts with descriptions, parameters, constraints, preconditions, relations, and utility evidence; structural filtering and ranking retrieve concepts for later image and video tasks.Full paper: §§3.3–3.4 · HTML full text · Figure 2

How memory changes (T2 operations)

  • Update · Prompted LLM/VLM record update

    Hard and soft visual feedback can add, update, or delete symbolic concepts and revise their parameters, constraints, and preconditions.

  • Consolidation · Merging & deduplication

    Memory optimization consolidates related symbolic concepts and resolves redundant entries.

  • Forgetting · Discrete eviction

    Concepts that remain invalid or unhelpful can be deprecated or deleted from the persistent repository.

How the policy evolves (T3)

  • Retrieval policy

    In sequential online groups, visual execution feedback revises persistent concept descriptions, preconditions, constraints, and relations only after each group; subsequent groups retrieve and compose the updated concepts.

Related systems