Multimodal Agent MemorySubmit a paper

EmbodiSkill

EmbodiSkill: Skill-Aware Reflection for Self-Evolving Embodied Agents

Ruofei Ju, Xinrui Wang, Xin Ding, Yifan Yang, Hao Wu, Shiqi Jiang, et al. · 2026-07-11

Read the paperPDF

Management
T1 Static Memory
Memory is built once and reused without change during deployment.
Primary modality
Embodied & 3D
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. evolved procedural skill document (skill body S_body plus skill appendix S_app)

    1. Represented asPerception-to-text records & summariesCross-represented explicit
    2. Stored inFlat sequential records
    3. Retrieved byUnclear
    4. Source evidenceSource not retained
    The persistent memory artifact is a single natural-language skill S^(n) = (S_body^(n), S_app^(n)), where S_body holds the prescriptive content reused in future task execution and S_app is an appendix marking which parts of the body are validated. It is a distilled textual record of prior execution experience, not retained observations: it specifies prerequisites, subgoal ordering, object affordances, visual-search strategies, action preconditions and recovery strategies. Raw trajectory frames are not stored, so the perceptual source is not recoverable from memory.Full paper: Section 3.1 (problem formulation)
    The skill reaches downstream control by being conditioned on directly in the executor policy a_t ~ pi_theta(. | I, S^(n), h_t), where h_t is the within-trajectory observation-action history. The executor parameters theta are kept fixed; all cross-task improvement is externalized into the skill text, so the skill document is the operative memory path.Full paper: Section 3.1 (executor definition)
    Skill content is produced by reflecting on executed trajectories from visual embodied environments: trajectories yield reflection records, records are consolidated, and skill-changing evidence rewrites S_body while execution-lapse evidence protects guidance that was valid but poorly executed. Visual-search strategies and object-affordance guidance in the skill derive from egocentric visual rollouts in EmbodiedBench-Habitat and EmbodiedBench-Navigation, so a non-text modality participates in memory formation.Full paper: Section 3.2.1 (skill-aware reflection), Section 3.2.2 (skill revision) · Figure 2

Related systems