Skip to content

Built an Agent Skill using Reflexion memory pattern — would love feedback #54

Description

@Thomaszhou22

Hi @noahshinn,

I built an installable Agent Skill that implements persistent self-reflection memory based on your Reflexion paper, combined with Self-Refine (Madaan et al.) as a plug-and-play skill for AI coding assistants.

What it does:

  • Core GENERATE→CRITIQUE→REFINE loop (from Self-Refine)
  • Cross-session Reflexion memory — your pattern of storing verbal self-reflections for future improvement
  • 6 reflection dimensions, 4 depth levels (auto-selected)
  • Works with Claude Code, Cursor, Copilot, Codex CLI, and 10+ more platforms
  • Zero API cost — host AI acts as its own critic

Benchmark: +21.2% average improvement across 20 test cases (5 categories).

Repo: https://github.com/Thomaszhou22/self-refine-skill

The Reflexion memory component stores error patterns discovered during self-review and carries them forward — exactly your approach of verbal reinforcement learning without parameter updates.

Would love your thoughts!

Best,
Thomas

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions