Multimodal Agent MemorySubmit a paper

ComfyClaw

ComfyClaw: Self-Evolving Skill Harnesses for Image Generation Workflows

Zongxia Li, Dawei Liu, Fuxiao Liu, Yuhang Zhou, Xiyang Wu, Jingxi Chen, et al. · 2026-07-02

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. versioned workflow skill library

    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
    The long-term skill library stores lightweight names, descriptions, tags, and trigger phrases separately from full step-by-step workflow procedures; a trigger router exposes relevant metadata first and retrieves full instructions on demand.Full paper: §§3.1–3.4 · HTML full text · Figure 1

How memory changes (T2 operations)

  • Consolidation · Experience abstraction

    Success and failure traces from completed image-generation runs are clustered and distilled into reusable workflow procedures rather than retained only as raw episodes.

  • Update · Prompted LLM/VLM record update

    The skill evolver applies CREATE, REVISE, and REINFORCE mutations to skill records, and accepted candidates are committed as new persistent versions.

  • Consolidation · Merging & deduplication

    The explicit MERGE mutation combines redundant skills before the candidate library is validated and committed.

  • Forgetting · Discrete eviction

    The explicit DELETE mutation removes a skill judged no longer useful, subject to the same validation and rollback process as other mutations.

How the policy evolves (T3)

  • Retrieval policy

    During deployed workflow runs, runtime errors and VLM feedback over rendered images drive validated REVISE and REINFORCE mutations that persistently change skill descriptions, triggers, and emphasis terms used by the trigger router; committed versions are invoked on later prompts.

Related systems