Multimodal Agent MemorySubmit a paper

SpaceMind

SpaceMind: A Modular and Self-Evolving Embodied Vision-Language Agent Framework for Autonomous On-orbit Servicing

Aodi Wu, Haodong Han, Xubo Luo, Ruisuo Wang, Shan He, Xue Wan · 2026-04-15

Read the paperPDF

Management
T3 Self-Evolving Memory
Feedback during deployment changes the memory-management policy itself.
Primary modality
Embodied & 3D
Also: Image, lidar, 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. persistent learned skill files for embodied servicing

    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
    Post-episode RGB, LiDAR, trajectory, and outcome evidence is distilled into learned skill files containing intent, triggers, rules, constraints, evidence, and provenance; the files are loaded for later episodes.Full paper: §3.5 · HTML full text · Algorithm 1

How memory changes (T2 operations)

  • Consolidation · Experience abstraction

    Complete multimodal servicing trajectories and outcomes are distilled into concise reusable skill rules.

  • Update · Prompted LLM/VLM record update

    The evolution module can create, overlay, rewrite, disable, or leave unchanged a learned skill file, subject to a quality gate.

How the policy evolves (T3)

  • Retrieval policy

    After each deployed episode, verified multimodal outcomes mutate persistent skill triggers, constraints, and rules; accepted skill files are loaded and selected during subsequent episodes.

Related systems