Multimodal Agent MemorySubmit a paper

MeMento

Long-Horizon Embodied Decision-Making via Multimodal Memory Compression

Bingxuan Li, Rui Yang, Cheng Qian, Jiateng Liu, Jeonghwan Kim, Zhenhailong Wang, et al. · 2026-08-02

Read the paperPDF

Management
T2 Dynamic Memory
Memory changes during deployment through a stable management policy.
Primary modality
Embodied & 3D
Also: Image, 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. preference-conditioned compressed multimodal history

    1. Represented asPersistent multimodal latent stateInternalized latent & parametric
    2. Stored inFlat sequential records
    3. Retrieved byLatent / parametric activation
    4. Source evidenceSource not retained
    A fixed number of memory tokens compress decision-relevant multimodal history conditioned on user preferences.Full paper: Multimodal Memory of Embodied Agent · PDF p. 2

How memory changes (T2 operations)

  • Update · Learned update module

    The trained compressor continually replaces long raw history with compact memory tokens. For MeMento, this specifically supports a trained but frozen update module of preference-conditioned compressed multimodal history.

  • Consolidation · Experience abstraction

    Long-horizon evidence is abstracted into preference-relevant decision state. For MeMento, this specifically supports abstraction of episodes into reusable experience of preference-conditioned compressed multimodal history.

Related systems