Multimodal Agent MemorySubmit a paper

ECHO

ECHO: Continuous Hierarchical Memory for Vision-Language-Action Models

Yanbin Hu, Jin Cui, Jiayi Lu, Ruixuan Yang, Jun Ye, Boran Zhao, et al. · 2026-05-09

Read the paperPDF

Management
T2 Dynamic Memory
Memory changes during deployment through a stable management policy.
Primary modality
Embodied & 3D
Also: text
Representation
Internalized latent & parametric

How experience becomes memory

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

  1. hyperbolic experience memory tree

    1. Represented asPersistent multimodal latent stateInternalized latent & parametric
    2. Stored inHierarchical multi-tier store
    3. Retrieved byStructure traversal, Latent / parametric activation
    4. Source evidenceSource not retained
    VLA hidden states are mapped into a continuous hyperbolic tree for top-down experience retrieval.Full paper: Autonomous Memory Consolidation · PDF p. 4 · Figure 1

How memory changes (T2 operations)

  • Update · Learned update module

    A fixed background mechanism inserts and refines experience vectors in continuous memory. For ECHO, this specifically supports a trained but frozen update module of hyperbolic experience memory tree.

  • Consolidation · Merging & deduplication

    During operation, ECHO applies merging or deduplication of overlapping entries to hyperbolic experience memory tree; the memory state changes while the governing update policy remains fixed.

  • Consolidation · Experience abstraction

    Background consolidation synthesizes virtual memories in the continuous space. For ECHO, this specifically supports abstraction of episodes into reusable experience of hyperbolic experience memory tree.

  • Reorganization · Relinking & reindexing

    Geometric interpolation and structural splitting reorganize the semantic memory tree. For ECHO, this specifically supports relinking or reindexing of stored items of hyperbolic experience memory tree.

Related systems