Multimodal Agent MemorySubmit a paper

TRAM

TRAM: Enhancing Multimodal Reasoning with Trajectory-Derived Auxiliary Memory

Kang Liu, Zijing Wang, Yongkang Liu, Mengjie Zhao, Xiaocui Yang, Shi Feng, et al. · 2026-08-03

Read the paperPDF

Management
T2 Dynamic Memory
Memory changes during deployment through a stable management policy.
Primary modality
Image
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. fast and slow trajectory latents

    1. Represented asPersistent multimodal latent stateInternalized latent & parametric
    2. Stored inFlat sequential records
    3. Retrieved byLatent / parametric activation
    4. Source evidenceSource not retained
    Completed reasoning is consolidated into fast and slow recurrent latent streams.Full paper: Latency (ms/token) Peak memory (GB) · PDF p. 7 · Figure 1

How memory changes (T2 operations)

  • Update · Learned update module

    Fixed recurrent modules update both streams during decoding. For TRAM, this specifically supports a trained but frozen update module of fast and slow trajectory latents.

Related systems