Multimodal Agent MemorySubmit a paper

EMKG

EMKG: Embodied Memory Knowledge Graphs for Object-Goal Navigation in Dynamic Open Worlds

Mingyi Li, Hui Liu, Ying Li, Shubo Zhang, Chunle Gao, Xiaokang Ma, et al. · 2026-01-01

Read the paper

Management
T2 Dynamic Memory
Memory changes during deployment through a stable management policy.
Primary modality
Embodied & 3D
Also: text, Image
Representation
Hybrid
Distinct operative representation paths are deliberately linked or jointly consumed by downstream reasoning, generation, or control.

How experience becomes memory

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

  1. Object nodes and target captions

    1. Represented asModality-specific encodings & geometric stateSource- and modality-specific
    2. Stored inGraph / relational store, Vector-indexed collection
    3. Retrieved bySimilarity search, Metadata / spatiotemporal retrieval, Structure traversal
    4. Source evidenceRetained but not accessible to the agent
    Builds an extensible multimodal memory relational knowledge graph.; Fuses image/text embeddings with 3D coordinate representations.; Sampling-based RAG update evaluates top-k candidates and incrementally updates the vector memory.. graph query/traversal over the memory conditioned on the current goal.Publisher method description: Dynamic Memory Knowledge Graph · Publisher pp. 4537–4544 · Figure 1 / method overview

  2. 3D coordinates and spatial relations

    1. Represented asPerception-to-text records & summariesCross-represented explicit
    2. Stored inGraph / relational store, Vector-indexed collection
    3. Retrieved bySimilarity search, Metadata / spatiotemporal retrieval, Structure traversal
    4. Source evidenceSource not retained
    Builds an extensible multimodal memory relational knowledge graph.; Fuses image/text embeddings with 3D coordinate representations.; Sampling-based RAG update evaluates top-k candidates and incrementally updates the vector memory.. graph query/traversal over the memory conditioned on the current goal.Publisher method description: Dynamic Memory Knowledge Graph · Publisher pp. 4537–4544 · Figure 1 / method overview

  3. Multimodal vector memory entries

    1. Represented asSpatial, temporal & cross-modal structured memoryCross-represented explicit
    2. Stored inGraph / relational store, Vector-indexed collection
    3. Retrieved bySimilarity search, Metadata / spatiotemporal retrieval, Structure traversal
    4. Source evidenceSource not retained
    Builds an extensible multimodal memory relational knowledge graph.; Fuses image/text embeddings with 3D coordinate representations.; Sampling-based RAG update evaluates top-k candidates and incrementally updates the vector memory.. graph query/traversal over the memory conditioned on the current goal.Publisher method description: Dynamic Memory Knowledge Graph · Publisher pp. 4537–4544 · Figure 1 / method overview

How memory changes (T2 operations)

  • Update · Rule-based update

    During operation, EMKG applies fixed-rule insertion or replacement to Object nodes and target captions; the memory state changes while the governing update policy remains fixed.

  • Revision · Evidence-weighted belief revision

    During operation, EMKG applies evidence-weighted revision of stored belief or confidence to Object nodes and target captions; the memory state changes while the governing update policy remains fixed.

Related systems