Multimodal Agent MemorySubmit a paper

GeoEvolver

Experience-Driven Multi-Agent Systems Are Training-free Context-aware Earth Observers

Pengyu Dai, Weihao Xuan, Junjue Wang, Hongruixuan Chen, Jian Song, Yafei Ou, et al. · 2026-01-30

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. global Earth-observation experience bank

    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
    The global bank persists distilled tool-chain patterns and failure-derived guardrails across queries and retrieves top-k entries by embedding similarity as strategy context.Full paper: §3.3 Hierarchical Memory · HTML full text · Equation 5

  2. episode-local compressed working context

    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
    Within an episode, older interactions are summarized into a compact progress state while a raw tail of recent tool steps remains available to the executors.Full paper: §3.3 Hierarchical Memory · HTML full text · Equation 6

How memory changes (T2 operations)

  • Consolidation · Experience abstraction

    Successful variants yield reusable operation order and decision checkpoints, while failed variants yield root-cause attributions and corrective guardrails; contrastive distillation extracts invariants across variants.

  • Consolidation · Merging & deduplication

    Single-variant and contrastive memories are consolidated into the global bank through canonical-key deduplication.

  • Consolidation · Episodic / hierarchical summarization

    A fixed working-memory operator summarizes older episode interactions while retaining a bounded recent trajectory tail.

How the policy evolves (T3)

  • Retrieval policy

    For each deployed Earth-observation query, the judge's success and validity signals plus tool errors drive distillation of reusable workflow templates and corrective guardrails; these persistent procedural priors are retrieved to alter orchestration and execution on future queries.

Related systems