Multimodal Agent MemorySubmit a paper

GeoForge

GeoForge: Non-Parametric Self-Evolving Agents for Earth-Observation Reasoning

Xin Xiao, Jiang Zhong, Junnan Zhu, Yingchao Feng, Peijin Wang, Yidan Zhang, et al. · 2026-08-11

Read the paperPDF

Management
T3 Self-Evolving Memory
Feedback during deployment changes the memory-management policy itself.
Primary modality
Image
Also: 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. reliability-aware workflow graph memory

    1. Represented asSpatial, temporal & cross-modal structured memoryCross-represented explicit
    2. Stored inGraph / relational store
    3. Retrieved byStructure traversal, Planner / tool-assisted retrieval
    4. Source evidenceSource not retained
    A reliability-aware directed graph stores Earth-observation workflow nodes, tool-transition edges, usage statistics, sensing dependencies, and operation order for task-conditioned traversal.Full paper: Workflow Graph Memory · HTML full text

  2. action-level execution experiences

    1. Represented asPerception-to-text records & summariesCross-represented explicit
    2. Stored inVector-indexed collection
    3. Retrieved bySimilarity search, Planner / tool-assisted retrieval
    4. Source evidenceSource not retained
    Compact action lessons preserve local tool corrections and are retrieved as part of the task-conditioned inference context.Full paper: Action-Level Experiences · HTML full text

  3. adapted skill standard operating procedure

    1. Represented asPerception-to-text records & summariesCross-represented explicit
    2. Stored inFlat sequential records
    3. Retrieved byPlanner / tool-assisted retrieval
    4. Source evidenceSource not retained
    A persistent adapted SOP records reusable procedures, data requirements, and parameter constraints and is concatenated with graph and action memories at inference.Full paper: Adapted Skill SOP and Evidence-Grounded Execution · HTML full text · Equation 13

How memory changes (T2 operations)

  • Consolidation · Experience abstraction

    After execution, a distiller converts the compacted grounded trajectory into reusable action lessons, a revised SOP, and workflow candidates while excluding sample-specific shortcuts.

  • Update · Prompted LLM/VLM record update

    The distiller proposes textual and structured memory edits, and a safety gate controls whether the revised SOP replaces the persistent prior.

  • Consolidation · Merging & deduplication

    Graph updates normalize and merge candidate workflows by tool-sequence and lexical similarity, while action experiences are inserted only after deduplication.

  • Forgetting · Discrete eviction

    The workflow-graph updater prunes stored structures using their accumulated reliability after merging new candidates.

How the policy evolves (T3)

  • Retrieval policy

    At inference, completed tool executions and their grounded outcomes generate a safety-gated revised SOP and workflow candidates; accepted artifacts persist in the nonparametric state and are retrieved to govern operation selection and composition on later Earth-observation tasks.

Related systems