Self Improving Agent CN

AI自我改进与记忆系统 - 解决'同类错误反复犯、用户纠正不长记性'的痛点。自动捕获错误、用户纠正、最佳实践，并转化为长期记忆。

Security Scan

The skill's code mostly matches a local memory/logger for corrections, errors, and best practices, but the SKILL.md overclaims automatic triggers and cross-project syncing and the scripts will store any supplied text (including secrets) without sanitization—review before use.

Purpose & Capability

The name/description promise an "automatic" self-improvement and cross-project memory system. The included Python scripts implement simple CLI logging and lookup to ~/.openclaw/memory/self-improving (errors.jsonl, corrections.jsonl, best_practices.jsonl) which is coherent with long-term memory storage, but there is no code to perform automatic detection of failing commands, listen for user utterances, sync to project .learnings/, or edit AGENTS.md/MEMORY.md despite the documentation claiming those features.

Instruction Scope

SKILL.md describes automatic triggers (on non-zero exit codes, specific user phrases) and automatic cross-project sync; runtime instructions primarily show manual CLI invocations of the provided scripts. The scripts only read/write local JSONL files and do not implement the automated monitoring, phrase-detection, or sync/edit operations described in the documentation.

Install Mechanism

No install spec and the package is instruction+scripts only; the Python files are included and only create/read files under the user's home (~/.openclaw/memory/self-improving). No network downloads or third-party package installs are performed by the skill itself.

Credentials

The skill requests no environment variables or external credentials. However, the scripts indiscriminately record whatever strings are passed to them (commands, errors, corrections), and there is no built-in sanitization or secret-detection—so they can persist sensitive data if invoked with such content.

Persistence & Privilege

The skill does not set always:true and does not modify other skills or system-wide configs. It persists data only under the user's home directory. Note: the platform-default that the agent may invoke the skill autonomously could cause the agent to record data automatically if configured to call these scripts, but that behavior is not implemented by the scripts themselves.

What to consider before installing

This skill is a local logger that writes JSONL records to ~/.openclaw/memory/self-improving and provides a simple check tool. Before installing or enabling it: (1) review the Python source yourself—there are no network calls but the scripts will persist any text you pass in; (2) avoid sending secrets/credentials to these log functions or add client-side redaction; (3) restrict file permissions on the memory directory (chmod 700) or encrypt the files if they may contain sensitive content; (4) be cautious about the SKILL.md claims — automatic detection, cross-project sync, and edits to AGENTS.md are not implemented in the provided code; if you need those features, request implementation details or an audited integration; (5) if you plan to allow autonomous agent invocation, ensure the agent's behaviors are constrained so it won't log sensitive context without explicit user consent. If additional information is provided showing network endpoints, automatic hooks, or sync code, re-evaluate (that would raise the severity).

Self-Improving Agent

让AI从错误中学习，越用越聪明。

核心解决的问题

✅ 命令/操作莫名失败，下次还用同样的错方式
✅ 反复纠正AI的写法、偏好、风格，它下个会话又忘了
✅ 同一个项目里反复踩同一个坑
✅ 发现更好的做法/更优解法，却没有系统化记住
✅ 外部工具/API变动，AI还在用旧知识
✅ 重要经验只存在于当前会话，跨天/跨项目就丢失

安装

# 创建记忆目录
mkdir -p ~/.openclaw/memory/self-improving

# 使用本skill
cat ~/.openclaw/skills/self-improving-agent/SKILL.md

使用方法

1. 记录错误（自动）

当命令失败时，自动记录到错误库：

# 脚本会自动捕获并记录
python3 ~/.openclaw/skills/self-improving-agent/log_error.py \
  --command "npm install xxx" \
  --error "permission denied" \
  --fix "use sudo or check permissions"

2. 记录用户纠正

当用户说：

• "不对，应该..."
• "错了，要用..."
• "不对，我之前说过..."

自动记录：

python3 ~/.openclaw/skills/self-improving-agent/log_correction.py \
  --topic "代码风格" \
  --wrong "用了双引号" \
  --correct "项目要求单引号" \
  --context "AGENTS.md 第23行"

3. 记录最佳实践

发现更好的做法时：

python3 ~/.openclaw/skills/self-improving-agent/log_best_practice.py \
  --category "security" \
  --practice "安装skill前必须审计代码" \
  --reason "防止供应链投毒"

4. 查看记忆（执行前自动检查）

# 执行命令前，先检查是否有相关记忆
python3 ~/.openclaw/skills/self-improving-agent/check_memory.py \
  --command "npm install"

输出示例：

⚠️ 发现相关记忆:
  [错误] 2026-03-05: npm install 权限失败 → 改用 sudo 或在用户目录安装
  [纠正] 2026-03-04: 项目使用单引号而非双引号

文件结构

~/.openclaw/memory/self-improving/
├── errors.jsonl          # 错误记录
├── corrections.jsonl     # 用户纠正
├── best_practices.jsonl  # 最佳实践
├── knowledge_gaps.jsonl  # 知识盲区
└── index.json           # 快速索引

自动触发场景

场景	自动记录	触发词
命令失败	✅ errors	退出码非0
用户纠正	✅ corrections	"不对" "错了" "应该"
更好的方法	✅ best_practices	"更好的" "更高效" "最优"
知识过时	✅ knowledge_gaps	"过时了" "已废弃" "新版是"

跨项目同步

重要记忆自动同步到：

• ~/.openclaw/memory/self-improving/ (全局)
• 当前项目 .learnings/ (项目级)
• AGENTS.md (用户偏好)
• MEMORY.md (长期记忆)

与现有Skill对比

功能	原self-improvement	本skill
记录错误	✅	✅
记录纠正	✅	✅ 增强版
自动触发	❌ 手动	✅ 自动
跨项目记忆	❌ 项目级	✅ 全局+项目
执行前检查	❌	✅
知识过时检测	❌	✅

示例工作流程

场景1：用户纠正代码风格

用户: "不对，我说过项目里要用单引号"

AI动作:

1. 自动记录纠正到 corrections.jsonl
2. 更新 AGENTS.md 添加代码风格规则
3. 回复: "已记录，下次会使用单引号"

场景2：命令反复失败

AI: 执行 npm install -g xxx 结果: 权限失败

AI动作:

1. 记录错误到 errors.jsonl
2. 下次执行前检查记忆
3. 自动改用: sudo npm install -g xxx 或本地安装

场景3：发现更好做法

用户: "其实用 pip install -e . 更高效"

AI动作:

1. 记录最佳实践到 best_practices.jsonl
2. 后续安装Python包时优先使用 editable 模式

注意事项

• 记忆文件定期备份到git
• 敏感信息脱敏后记录
• 定期review并清理过时记忆

---

Created: 2026-03-05 by 老二