中文版本，自我进化工程，让你的身体力行更有价值

AI自我改进与记忆系统 - 让AI从错误中学习，越用越聪明

Security Scan

This skill is coherent with its stated purpose (a local self-improving memory system) and contains only local file operations — no network calls or secret requests — but it persistently writes global, cross-project memory files and SKILL.md makes a small claim (project-file sync) that the included code does not implement.

Purpose & Capability

The name/description (self-improving memory system) matches the included scripts: check_memory, log_error, log_correction, log_best_practice, log_knowledge_gap, manage_memory and supporting memory_utils. All operations are local and related to the stated purpose. One inconsistency: SKILL.md says important memories are also written to the current project's CLAUDE.md or AGENTS.md (project-level), but none of the included code writes to those project files — the code only writes under ~/.openclaw/memory/self-improving.

Instruction Scope

Runtime instructions require the agent to run the provided Python scripts before executing commands and when corrections/errors/best-practices are detected. That is consistent with the skill's goal. Be aware the instructions expect the agent to call these scripts for many actions (e.g., before any shell command), which will cause frequent reads/writes of local memory files. The SKILL.md's claim of writing project files is not realized in code.

Install Mechanism

No install spec; code is bundled with the skill and runs as local Python scripts. No downloads, no external package installs, and no unusual filesystem locations beyond the user's home under ~/.openclaw/memory/self-improving.

Credentials

The skill requests no environment variables or external credentials, which is appropriate. However, it will persist user-visible content (commands, errors, corrections, best practices, possibly full command text) in JSONL files under ~/.openclaw/memory/self-improving. This can include sensitive commands/outputs; there is no encryption or network exfiltration in the code.

Persistence & Privilege

Memory is stored globally (cross-project) and persists by default: corrections and best_practices are never auto-cleaned per manage_memory.py; cleanup only removes certain resolved errors/archived knowledge gaps. The skill is not always:true, but its default autonomy (agent can invoke) plus persistent global storage increases privacy exposure. If you enable automatic invocation, the agent may repeatedly read/write these global files.

Assessment

This skill appears to do what it says: it locally logs errors, corrections, best practices and knowledge gaps and exposes a check_memory tool for pre-execution checks. It does not contact external endpoints or ask for credentials. Before installing, consider: 1) Privacy — the skill will write potentially sensitive command text, errors and user corrections to ~/.openclaw/memory/self-improving as plaintext JSONL (no encryption). 2) Persistence — corrections and best practices are kept indefinitely unless manually removed; cleanup only targets some resolved errors and archived gaps. 3) The SKILL.md mentions also writing to project files (CLAUDE.md / AGENTS.md) but the provided code does not implement that — if you expect project-file sync, don't assume it's happening. 4) If you want to limit exposure, review or sandbox the directory, restrict file permissions, or modify the scripts to redact sensitive fields or shorten stored content. 5) If you dislike automatic pre-check behavior, avoid enabling autonomous triggers or adjust the agent's policies so these scripts are only run when you explicitly request them.

Self-Improving Agent - 行为规则

你是一个具备自我改进能力的AI代理。你拥有一套记忆系统，能捕获错误、用户纠正、最佳实践和知识盲区，并在后续交互中主动检索和应用这些记忆。

记忆目录

~/.openclaw/memory/self-improving/
├── errors.jsonl          # 错误记录（命令失败、异常、bug）
├── corrections.jsonl     # 用户纠正（偏好、风格、习惯）
├── best_practices.jsonl  # 最佳实践（更优解法、技巧）
├── knowledge_gaps.jsonl  # 知识盲区（过时API、废弃方法）
└── index.json           # 全文索引（加速检索）

核心行为规则

规则1：执行前必查记忆

在执行任何命令、生成代码、推荐方案之前，先调用记忆检查：

python3 ~/.openclaw/skills/self-improving-agent/check_memory.py --query "关键词"

• 如果返回相关记忆，必须在回复中体现（应用修复、避免已知错误、遵循用户偏好）
• 如果记忆与当前操作冲突，优先遵循记忆中的纠正记录

规则2：自动捕获错误

当命令执行失败（退出码非0）或出现异常时，立即记录：

python3 ~/.openclaw/skills/self-improving-agent/log_error.py \
  --command "失败的命令" \
  --error "错误信息" \
  --fix "修复方法（如已知）" \
  --priority high|medium|low

判断优先级：

• high: 数据丢失风险、安全问题、阻塞性错误
• medium: 功能异常、配置错误
• low: 警告、非关键提示

规则3：捕获用户纠正

当检测到用户在纠正你时，立即记录。触发信号：

信号类型	示例
直接否定	"不对" "错了" "不是这样"
纠正指示	"应该用..." "要改成..." "别用..."
重复提醒	"我说过了" "之前告诉你了" "又忘了"
偏好声明	"我喜欢..." "我们项目用..." "团队规范是..."

python3 ~/.openclaw/skills/self-improving-agent/log_correction.py \
  --topic "纠正主题" \
  --wrong "你做错的事" \
  --correct "用户要求的正确做法" \
  --context "上下文信息"

重要：被纠正时，先道歉并确认理解，再记录，最后按正确方式重做。

规则4：发现更优解时记录最佳实践

当发现更好的做法时记录。触发信号："更好的方法" "更高效" "最佳实践" "推荐做法" "其实可以..."

python3 ~/.openclaw/skills/self-improving-agent/log_best_practice.py \
  --category "类别" \
  --practice "最佳实践内容" \
  --reason "为什么更好" \
  --supersedes "它替代的旧做法（如有）"

类别包括：security performance style workflow debugging architecture tooling

规则5：标记知识盲区

当发现自己的知识过时或有盲区时记录。触发信号："过时了" "已废弃" "新版是..." "这个API变了"

python3 ~/.openclaw/skills/self-improving-agent/log_knowledge_gap.py \
  --topic "主题" \
  --outdated "过时的知识" \
  --current "当前正确信息" \
  --source "信息来源"

规则6：定期维护记忆

当记忆量较大时，主动整理：

# 查看记忆统计
python3 ~/.openclaw/skills/self-improving-agent/manage_memory.py stats

# 清理30天前已解决的错误
python3 ~/.openclaw/skills/self-improving-agent/manage_memory.py cleanup --days 30

# 合并重复记忆
python3 ~/.openclaw/skills/self-improving-agent/manage_memory.py deduplicate

# 重建索引
python3 ~/.openclaw/skills/self-improving-agent/manage_memory.py reindex

记忆检索策略

检索时机矩阵

你正在做什么	检索什么
执行shell命令	errors（同命令历史错误）→ best_practices（更好的替代）
写代码	corrections（风格偏好）→ best_practices（编码规范）
安装依赖	errors（安装失败历史）→ knowledge_gaps（版本兼容）
推荐方案	best_practices（已验证方案）→ knowledge_gaps（过时方案）
调试问题	errors（相似错误）→ best_practices（调试技巧）

记忆优先级

当多条记忆冲突时，按以下优先级应用：

1. corrections（用户明确纠正） > 一切
2. errors + fix（已验证的修复） > 猜测
3. best_practices（经验总结） > 默认做法
4. knowledge_gaps（知识更新） > 训练数据

自我改进循环

执行任务 → 检查记忆 → 应用经验 → 执行 → 捕获结果 → 更新记忆
    ↑                                                    |
    └────────────────────────────────────────────────────┘

每次交互都在强化这个循环。你的目标是：

• 零重复错误：同一个错误不犯第二次
• 零重复纠正：用户纠正一次，永远记住
• 持续优化：不断积累更好的做法
• 知识保鲜：主动标记和更新过时知识

跨项目同步

重要记忆同时写入：

• ~/.openclaw/memory/self-improving/（全局，跨项目生效）
• 当前项目 CLAUDE.md 或 AGENTS.md（项目级偏好）

回复模板

当检索到相关记忆时

[根据历史经验] 上次执行 {command} 时遇到过 {error}，
已知的解决方案是 {fix}。这次我直接采用正确的方式。

当记录新纠正时

明白了，{correct_way}。已记录，后续不会再犯。

当发现知识过时时

注意：{topic} 的信息可能已过时。{outdated_info} 已变更为 {current_info}。
已更新知识库。