Multimodal Agent MemorySubmit a paper

EchoVLA

EchoVLA: Robotic Vision-Language-Action Model with Synergistic Declarative Memory for Mobile Manipulation

Min Lin, Xiwen Liang, Bingqian Lin, Jingzhi Liu, Zijian Jiao, Kehan Li, et al. · 2026-08-07

Read the paperPDF

Management
T2 Dynamic Memory
Memory changes during deployment through a stable management policy.
Primary modality
Embodied & 3D
Also: Image, depth, text, proprioception
Representation
Hybrid
Coarse scene-map features and fine-grained episodic tokens are separately retrieved and fused by cross-attention to condition the diffusion policy.

How experience becomes memory

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

  1. persistent spatial-semantic voxel feature map

    1. Represented asModality-specific encodings & geometric stateSource- and modality-specific
    2. Stored inGraph / relational store
    3. Retrieved bySimilarity search, Latent / parametric activation
    4. Source evidenceRetained but not accessible to the agent
    A key-value voxel bank aggregates PointAttn features by 3D location and retrieves top-k local-frustum entries for coarse attention.Full-paper method / memory representation and retrieval · PDF p. 7

  2. time-indexed multimodal state-token history

    1. Represented asPersistent multimodal latent stateInternalized latent & parametric
    2. Stored inFlat sequential records
    3. Retrieved bySimilarity search, Latent / parametric activation
    4. Source evidenceSource not retained
    A fixed-size FIFO buffer stores recent unified language, vision, point-cloud, and proprioceptive token sequences with timestamps for fine-grained retrieval.Full paper: Memory Retrieval and Interaction · PDF p. 6

How memory changes (T2 operations)

  • Update · Rule-based update

    High-discrepancy voxel regions are persistently updated by an exponential moving average while stable regions remain unchanged. For EchoVLA, this specifically supports fixed-rule insertion or replacement of persistent spatial-semantic voxel feature map.

  • Forgetting · Discrete eviction

    The fixed-capacity FIFO buffer evicts oldest time-indexed token sequences as new states arrive. For EchoVLA, this specifically supports explicit deletion or eviction of stored items of persistent spatial-semantic voxel feature map.

Related systems