Multimodal Agent MemorySubmit a paper

Chain-of-Memory

Chain-of-Memory: Enhancing GUI Agents for Cross-Application Navigation

Xinzge Gao, Chuanrui Hu, Bin Chen, Teng Li · 2025-06-22

Read the paperPDF

Management
T2 Dynamic Memory
Memory changes during deployment through a stable management policy.
Primary modality
GUI & computer use
Also: Image, text
Representation
Cross-represented explicit

How experience becomes memory

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

  1. recent operation descriptions and persistent task-relevant screen information

    1. Represented asPerception-to-text records & summariesCross-represented explicit
    2. Stored inHierarchical multi-tier store
    3. Retrieved byPlanner / tool-assisted retrieval
    4. Source evidenceSource not retained
    Information perception summarizes prior screens and actions; STM tracks recent task state and LTM preserves key screen information for action decisions.Full paper: Memory · PDF p. 2 · Figure 2

How memory changes (T2 operations)

  • Update · Prompted LLM/VLM record update

    At every step the model updates STM and decides which extracted screen facts should be added to persistent LTM. For Chain-of-Memory, this specifically supports LLM-mediated memory editing of recent operation descriptions and persistent task-relevant screen information.

  • Consolidation · Episodic / hierarchical summarization

    During operation, Chain-of-Memory applies episodic or hierarchical summarization to recent operation descriptions and persistent task-relevant screen information; the memory state changes while the governing update policy remains fixed.

  • Forgetting · Discrete eviction

    During operation, Chain-of-Memory applies explicit deletion or eviction of stored items to recent operation descriptions and persistent task-relevant screen information; the memory state changes while the governing update policy remains fixed.

Related systems