reflexion（reflexion 搜索）

Reflexion

Closed-loop learning for AI coding agents. Inspired by Reflexion: Language Agents with Verbal Reinforcement Learning.

*"Reflexion agents verbally reflect on task feedback signals, then maintain their own reflective text in an episodic memory buffer to induce better decision-making in subsequent trials."* — Shinn et al., 2023

The problem: AI agents repeat the same mistakes across sessions. They don't learn from errors, don't remember corrections, and every new conversation starts from zero.

The solution: A capture-recall-promote loop that closes the feedback gap.

Error/Correction occurs
        |
   [CAPTURE] -----> .reflexion/entries/
        |                    |
   Next similar task    [INDEX] keywords
        |                    |
   [RECALL] <--- keyword match on prompt
        |
   Inject past solution into context
        |
   [VERIFY] did it work?
        |           |
      Yes          No ---> update entry, flag for review
        |
   occurrences >= 3?
        |           |
      Yes          No ---> increment counter
        |
   [PROMOTE] append rule to CLAUDE.md

Quick Reference

Install

Claude Code (recommended)

# Clone into your project or global skills
git clone https://github.com/user/reflexion.git .claude/skills/reflexion

# Or copy into an existing skills directory
cp -r reflexion/ ~/.claude/skills/reflexion

Add hooks to .claude/settings.json:

{
  "hooks": {
    "PostToolUse": [
      {
        "matcher": "Bash",
        "hooks": [
          {
            "type": "command",
            "command": "./.claude/skills/reflexion/scripts/capture.sh"
          }
        ]
      }
    ],
    "UserPromptSubmit": [
      {
        "matcher": "",
        "hooks": [
          {
            "type": "command",
            "command": "./.claude/skills/reflexion/scripts/recall.sh"
          }
        ]
      }
    ]
  }
}

First run

The scripts auto-initialize on first use. No setup needed. To manually initialize:

./scripts/init.sh

How It Works

1. Capture (automatic)

The capture.sh hook fires after every Bash tool use. It reads the tool output from stdin (JSON), detects errors via pattern matching, and stores structured entries:

{
  "id": "RFX-20260331-a7f",
  "type": "error",
  "trigger": "npm ERR! Missing script: \"build\"",
  "context": "npm run build",
  "resolution": "",
  "keywords": ["npm", "build", "missing", "script"],
  "occurrences": 1,
  "first_seen": "2026-03-31",
  "last_seen": "2026-03-31",
  "promoted": false,
  "cwd": "/home/user/project"
}

When the agent (or user) resolves the error, the agent should update the entry:

Update .reflexion/entries/RFX-20260331-a7f.json with resolution:
"Use pnpm run build - this project uses pnpm, not npm"

2. Recall (automatic)

The recall.sh hook fires before every user prompt. It extracts keywords from the prompt, searches the entry index, and injects relevant past learnings:

<reflexion-recall>
Past learning [RFX-20260331-a7f] (seen 2x):
  Trigger: npm ERR! Missing script: "build"
  Resolution: Use pnpm run build - this project uses pnpm, not npm
  Keywords: npm, build, missing, script
</reflexion-recall>

This costs ~50-80 tokens when matches exist, zero when they don't.

3. Promote (automatic)

When an entry hits 3+ occurrences, promote.sh appends a concise rule to CLAUDE.md:

<!-- reflexion:auto-promoted -->
## Reflexion: Learned Rules

- This project uses pnpm, not npm. Always use `pnpm run` commands. (seen 3x, source: RFX-20260331-a7f)

Promoted entries are marked "promoted": true and stop being injected via recall (the rule is now in CLAUDE.md permanently).

4. Verify (agent-driven)

After the agent applies a recalled solution, it should verify and update:

• Worked: Increment occurrences, update last_seen
• Failed: Add note to entry, flag for review, decrement confidence

This step is agent-driven (via prompt instruction), not hook-automated, to avoid false positives.

Entry Types

Data Format

Entries live in .reflexion/entries/ as individual JSON files (one per learning). This enables:

• Fast grep-based search (no parsing a giant markdown file)
• Atomic writes (no corruption from concurrent access)
• Easy manual editing
• Git-friendly diffs

The keyword index at .reflexion/index.txt maps keywords to entry IDs for fast recall:

npm:RFX-20260331-a7f,RFX-20260401-b2c
build:RFX-20260331-a7f
pnpm:RFX-20260331-a7f,RFX-20260401-b2c
docker:RFX-20260402-c1d

Promotion Rules

An entry is auto-promoted to CLAUDE.md when ALL conditions are met:

1. occurrences >= 3
2. resolution is non-empty (the fix is known)
3. promoted is false (not already promoted)
4. Entry is older than 1 day (not a flurry of the same error in one session)

Promoted rules are written as short, actionable directives. Not incident reports.

Agent Instructions

When this skill is active, follow these behaviors:

On Error

1. Check if capture.sh already logged it (it runs automatically on Bash errors)
2. If you resolve the error, update the entry's resolution field
3. If the error matches a recalled learning, say so and apply the known fix

On User Correction

Log a correction entry manually:

cat > .reflexion/entries/RFX-$(date +%Y%m%d)-$(head -c3 /dev/urandom | xxd -p | head -c3).json << 'ENTRY'
{
  "id": "RFX-...",
  "type": "correction",
  "trigger": "user said: actually use pnpm",
  "context": "attempted npm install",
  "resolution": "this project uses pnpm, not npm",
  "keywords": ["npm", "pnpm", "install", "package-manager"],
  "occurrences": 1,
  "first_seen": "2026-03-31",
  "last_seen": "2026-03-31",
  "promoted": false
}
ENTRY

Then rebuild the index: ./scripts/rebuild-index.sh

On Recall

When <reflexion-recall> context appears in the prompt:

1. Read the recalled learnings
2. Apply the known resolution if relevant
3. If the resolution works, increment occurrences
4. If it doesn't apply, ignore it (no penalty)

Before Major Tasks

Run ./scripts/status.sh to see if there are relevant learnings for the area you're about to work in.

Security

• Never log secrets, tokens, API keys, or credentials in entries
• The capture.sh script redacts common secret patterns (Bearer tokens, API keys, passwords)
• .reflexion/ should be in .gitignore for private projects
• For team projects, committing .reflexion/ creates shared learning (opt-in)

Comparison

File Structure

reflexion/
├── SKILL.md                 # This file
├── scripts/
│   ├── init.sh              # Initialize .reflexion/ directory
│   ├── capture.sh           # PostToolUse hook - auto-capture errors
│   ├── recall.sh            # UserPromptSubmit hook - inject past learnings
│   ├── promote.sh           # Auto-promote recurring patterns to CLAUDE.md
│   ├── status.sh            # Learning stats dashboard
│   └── rebuild-index.sh     # Rebuild keyword index from entries
├── assets/
│   └── settings-template.json  # Claude Code settings template
└── references/
    └── integration.md       # Setup guides for different agents

Citation

This skill implements the core feedback loop from:

@article{shinn2023reflexion,
  title   = {Reflexion: Language Agents with Verbal Reinforcement Learning},
  author  = {Noah Shinn and Federico Cassano and Edward Berman and
             Ashwin Gopinath and Karthik Narasimhan and Shunyu Yao},
  journal = {arXiv preprint arXiv:2303.11366},
  year    = {2023},
  url     = {https://arxiv.org/abs/2303.11366},
  doi     = {10.48550/arXiv.2303.11366}
}

The paper showed that language agents reflecting on past failures in an episodic memory buffer significantly outperform base agents — achieving 91% pass@1 on HumanEval vs GPT-4's 80%. This skill adapts that principle for AI coding agents: instead of weight updates, it stores verbal reflections (error entries with resolutions) and retrieves them when similar situations arise.