Continuous Learning

持续学习套件 - AI自主记忆管理工作流：自动识别新任务、记录对话到MemPalace、定期做梦分析、提取精华到文档、自我纠错改进。触发词："持续学习"、"记忆管理"、"自我改进"、"学习体系"。

Security Scan

Skill's functionality mostly matches its description, but there are several inconsistencies and persistence behaviors (hardcoded absolute paths, scheduled-task installation, and implicit handling of API keys/notifications) that warrant review before installing.

Purpose & Capability

The skill claims to be a MemPalace-driven continuous-learning kit and the code implements that (reading memory entries, extracting summaries, updating documents). However, the code hardcodes absolute workspace and database paths (e.g., 'C:/Users/18625/.openclaw/workspace', 'C:/Users/18625/.mempalace/palace'), which is environment-specific and unexpected in a reusable skill. The use of scheduled tasks and cross-skill reads (skills/<name>/SKILL.md) are consistent with persistent memory management but raise proportionality questions: a simple 'learning helper' would not normally hardcode another user's path nor assume those exact locations.

Instruction Scope

SKILL.md and bootstrap_rules explicitly instruct the agent to read a wide set of local files (SOUL.md, AGENTS.md, MEMORY.md, TOOLS.md, .learnings/*, memory/YYYY-MM-DD.md) and even other skills' SKILL.md. The code implements reading local DB and files and queueing notifications. While this is necessary for the stated purpose, the instructions are broad (e.g., '任何不确定的情况' → treat as new task) and could cause the agent to access many user documents automatically. The scripts also call into a mempalace RPC (handle_request) which may trigger further behavior in that package—users need to audit mempalace's behavior as well.

Install Mechanism

There is no external download/install spec in the registry entry (instruction-only), and included Python scripts and a requirements.txt are provided. That lowers remote-code-install risk. The provided setup/install_cron.py does modify system scheduled tasks (cron / Windows Task Scheduler) which is a local, explicit installation step requiring user confirmation; this is expected for a periodic background job but is a persistence action to be aware of.

Credentials

The skill declares no required environment variables, but dream_config.json contains an 'api_key' for the MiniMax API and the code loads configuration from a local file (MINIMAX_CONFIG). Credentials are expected to be stored in config files rather than declared environment variables; this is not inherently malicious but is inconsistent with registry metadata (no creds declared) and makes secret handling less explicit. The skill also expects and writes files in user workspace paths, and will create/append to documents (.learnings, SOUL.md, AGENTS.md, etc.). Notification configuration contains a 'to' field that could direct messages externally (e.g., WeChat); although the code only writes to a local queue, that implies an external sender elsewhere—users should verify what actually sends queued notifications. Overall the level of file and credential access is high but largely explained by the feature set; the hardcoded absolute paths are a disproportionate implementation choice.

Persistence & Privilege

The skill installs persistent scheduled tasks (cron or Windows scheduled tasks) to run synchronization and 'dream' cycles at fixed times. always:false, so it won't force-enable itself, but the install script will add persistent jobs to the host when run (it prompts for confirmation). Persistent scheduling is reasonable for a daily 'dream' job, but increases blast radius if combined with any unintended data-exfiltration or network calls — verify what the scheduled scripts do and audit notification delivery components.

What to consider before installing

Key things to check before installing/activating: 1. Hardcoded paths: The scripts use absolute Windows paths (C:/Users/18625/...). Edit the scripts (or set a proper WORKSPACE) so they point to your workspace; otherwise the skill may fail silently or read/write unexpected locations. 2. Configured API key: MiniMax API key is expected in config/dream_config.json (analysis.api_key). Do not put secrets in world-readable files. Consider using a secure store or environment variables and update the code to read them. 3. Audit MemPalace: The sync script imports and calls mempalace.mcp_server.handle_request. Review the mempalace skill/package to understand what that RPC does and whether it communicates externally or handles sensitive data. 4. Scheduled tasks: install_cron.py will create persistent cron/Task Scheduler jobs. Only run it if you want background daily jobs. Inspect the exact cron entries / scheduled-task commands before applying; back up your crontab if needed. 5. Notification pipeline: The code writes messages into a local notification queue file; find and audit the component that reads that queue and actually sends WeChat/email messages to ensure no unexpected outbound communications. 6. Least privilege & sandboxing: If you want to evaluate the skill first, run it in a sandbox or throwaway workspace, and do not supply real API keys or production data until you’re comfortable with behavior. 7. Replace or parameterize workspace: Prefer using relative or user-configurable workspace paths rather than hardcoded ones; ensure file permissions are appropriate and the skill cannot read unrelated sensitive files. If you want, I can point to the exact lines to change for the path/config handling, or produce a checklist/patch to make the skill safer to test (e.g., replace hardcoded WORKSPACE with environment or config lookup).

持续学习套件 (Continuous Learning Kit)

让AI具备持续学习和自我进化的核心能力。

---

🧠 核心概念

工作流程

每条消息 → 新任务判断 → 记录到MemPalace → 每日定时做梦 → 提取精华到文档 → 自我改进

                ↓
         是否是新任务？
           Yes → 读技能文档 → 执行
           No  → 继续

两大周期

短期周期（每次消息）：

1. 新任务判断
2. 自动记录对话到MemPalace

长期周期（每日定时）：

1. 做梦分析（23:00同步聊天）
2. 精华提取（02:00做梦分析）
3. 文档更新
4. 自我纠错（ERRORS.md/LEARNINGS.md）

---

📦 技能包结构

continuous-learning/
├── SKILL.md                          # 技能说明（本文件）
├── bootstraps/
│   └── bootstrap_rules.md            # 启动规则（整合版BOOTSTRAP）
├── sync/
│   ├── sync_chats_daily.py           # 每日同步脚本
│   └── sync_notification.py          # 带通知的同步
├── dream/
│   ├── dream_cycle.py                # 做梦分析核心
│   ├── dream_notification.py         # 带通知版本
│   └── prompts/
│       ├── analysis_prompt.txt       # MiniMax分析提示词
│       └── extraction_rules.md       # 提取规则
├── notifications/
│   ├── notification_queue.json       # 通知队列（动态生成）
│   └── send_notifications.py         # 通知发送
├── config/
│   ├── dream_config.json             # 做梦配置
│   └── documentation_targets.json    # 文档目标
└── setup/
    ├── install_cron.py               # 定时任务安装（跨平台）
    └── init_learning_files.py        # 初始化学习文件

---

🎯 核心特性

1. 新任务智能判断

判断标准：

• 话题跨度大（从A项目跳到B项目）
• 任务类型变（查LIMS → 发邮件）
• 关键词第一次出现
• 用户说"新任务"、"开始做..."等

触发流程：

# 伪代码：新任务判断
if is_new_task(message, context):
    # 确认后读取：
    read("SOUL.md")
    read("AGENTS.md")
    read("MEMORY.md")
    read("TOOLS.md")

    # 读取学习文件
    read(".learnings/ERRORS.md")
    read(".learnings/LEARNINGS.md")

    # 今日记忆
    read(f"memory/{today}.md")
    read(f"memory/{yesterday}.md")

2. MemPalace自动记录

记录时机：

• 每日23:00定时
• 手动触发

记录内容：

SESSION:YYYY-MM-DD
| 对话摘要
| 用户偏好
| 项目背景
| 重要决策
| 错误教训

通知机制：

• 同步完成 → WeChat通知
• 做梦完成 → WeChat通知

3. 做梦分析

执行时间：每日02:00

分析流程：

1. 读取MemPalace所有碎片记忆（N条）
2. MiniMax M2.7分析→ 分类+价值判断+去重
3. 提取精华到5个核心文档：
   - SOUL.md     → 个性、偏好、风格
   - AGENTS.md   → 工作流程、规则
   - MEMORY.md   → 用户背景、项目
   - TOOLS.md    → 配置、坑、技巧
   - BOOTSTRAP.md → 启动规则
4. 生成分析报告
5. WeChat通知用户

4. 自我纠错

错误记录（ERRORS.md）：

### 错误：API字段猜测
**错误**: 猜测LIMS API用sampleBaseUuid获取报告
**正确**: 应该用sampleBaseTestingUuid
**教训**: 先查证，不要猜测
**日期**: 2026-04-10

学习记录（LEARNINGS.md）：

### 学习：握手流程
**收获**: 登录类API必须先GET再POST
**应用**: 企业微信、LIMS登录都适用
**日期**: 2026-04-10

---

⚠️ 前置条件（必需）

本技能包严格依赖MemPalace记忆系统，使用前必须先安装。

安装MemPalace

方法1：通过ClawHub安装（推荐）

clawdhub install mempalace

方法2：手动安装

下载技能包到你的技能目录：

cd skills/mempalace
# 确保以下文件存在：
# - SKILL.md
# - scripts/mcp_server.py
# - scripts/call.py

验证MemPalace

测试连接：

python skills/mempalace/scripts/call.py mempalace_status

期望输出：

{
  "status": "ready",
  "drawer_count": 1,
  "wings": ["你的wing名称"]
}

如果返回错误，请检查：

1. ChromaDB是否已安装：pip install chromadb
2. 数据库路径是否有写入权限

---

🚀 快速开始

步骤1：初始化学习文件

python setup/init_learning_files.py

创建：

• .learnings/ERRORS.md
• .learnings/LEARNINGS.md
• .learnings/FEATURES.md
• memory/YYYY-MM-DD.md

步骤2：配置文档目标

编辑 config/documentation_targets.json:

{
  "SOUL": {
    "path": "SOUL.md",
    "purpose": "行为准则、个性偏好、沟通风格"
  },
  "AGENTS": {
    "path": "AGENTS.md",
    "purpose": "工作流程、代理规则、交互模式"
  },
  "MEMORY": {
    "path": "MEMORY.md",
    "purpose": "用户偏好、项目背景、长期记忆"
  },
  "TOOLS": {
    "path": "TOOLS.md",
    "purpose": "工具配置、集成注意事项、坑"
  },
  "BOOTSTRAP": {
    "path": "BOOTSTRAP.md",
    "purpose": "会话启动规则、新任务判断"
  }
}

步骤3：安装定时任务

所有平台（Windows/Linux/Mac）:

python setup/install_cron.py

Windows用户注意：

• 如果提示权限问题，以管理员身份运行PowerShell/CMD
• 脚本会自动检测操作系统并配置相应的定时任务

步骤4：配置MiniMax API（可选）

编辑 config/dream_config.json:

{
  "analysis_model": {
    "provider": "minimax",
    "model": "MiniMax-M2.7",
    "api_url": "https://api.minimax.chat/v1",
    "api_key": "your_api_key_here"
  }
}

注意：如果不配置，会跳过大模型分析，只做基础分类。

---

📋 使用场景

场景1：多项目并行工作

用户：

查一下LIMS样本SDAA25D03362的报告

帮我做康鑫达周报

写一个Python脚本处理Excel

技能行为：

1. 第1条 → 检测到"LIMS"关键词 → 判断为新任务
2. 读取AGENTS.md → 知道LIMS地址和账号
3. 执行查询 → 记录结果到MemPalace
4. 第2条 → 检测到"康鑫达周报" → 新任务
5. 读取WORKFLOW.md → 调用周报生成脚本

场景2：持续优化技能

第1天：

• 执行任务A
• 失败 → 记录到ERRORS.md

第23:00：

• 同步今日对话到MemPalace

第02:00：

• 做梦分析 → 发现失败模式
• 提取教训到LEARNINGS.md
• 找到原因："API猜测"

第3天：

• 类似任务B出现
• 读取LEARNINGS.md
• 避免API猜测 → 成功！

场景3：用户偏好记忆

用户说："叫我xiaolong，不要叫刘总"

技能行为：

1. 记录到MemPalace
2. 做梦分析
3. 提取到SOUL.md：称呼偏好
4. 后续所有对话 → 直接用"xiaolong"

---

📊 配置选项

做梦配置 (config/dream_config.json)

{
  "sync_schedule": "23:00",
  "dream_schedule": "02:00",
  "notification_enabled": true,
  "notification_channel": "openclaw-weixin",
  "doc_update_rules": {
    "min_similarity": 0.7,
    "max_noise_ratio": 0.5
  }
}

分析模型配置

支持多个模型：

• MiniMax M2.7（默认，效果最好）
• 智谱GLM-4.7
• Claude 3.5 Sonnet
• GPT-4o

---

🔧 故障排查

问题1：定时任务不执行

Windows：

schtasks /query /tn "OpenClaw-SyncWeChat"
schtasks /query /tn "OpenClaw-DreamCycle"

检查状态是否为"就绪"。

Linux：

crontab -l | grep openclaw

问题2：MemPalace写入失败

检查：

1. ChromaDB路径是否正确
2. 写入权限是否足够
3. 数据库大小（超过5GB需清理）

问题3：做梦分析卡住

检查：

1. API Key是否有效
2. 网络连接
3. 记忆条目数（超过1000条需分批）

---

📈 进阶用法

自定义文档目标

添加自己的文档：

{
  "PROJECT_NOTES": {
    "path": "docs/project_notes.md",
    "purpose": "项目特定笔记"
  },
  "API_REFERENCE": {
    "path": "docs/api_reference.md",
    "purpose": "API调用记录"
  }
}

自定义分析提示词

编辑 dream/prompts/analysis_prompt.txt，改变分析规则。

多Agent协同

多个AI共享MemPalace：

{
  "shared_agents": ["agent_A", "agent_B"],
  "sync_interval": "hourly"
}

---

🤝 依赖技能

必需

• mempalace - 记忆存储系统

可选

• minimax-image-understanding - 图像记忆
• self-improving-agent - 互补的自我改进功能

---

📝 版本历史

v1.0.0 (2026-04-19)

• ✅ 新任务智能判断
• ✅ MemPalace自动记录
• ✅ 做梦分析（MiniMax M2.7）
• ✅ 5文档自动更新
• ✅ WeChat通知
• ✅ 自我纠错（ERRORS/LEARNINGS）

---

🌟 核心价值

这个技能包让AI从"一次性工具"进化为"持续学习的智能体"：

维度	传统AI	持续学习AI
记忆	每次对话重新开始	跨会话持久记忆
学习	需要人工提示	自动提取规律
纠错	重复犯同样错误	从错误中学习
偏好	不知道用户喜好	记住用户习惯
进化	能力固定	持续自我提升

---

技能作者: 小麦 (Xiaomai) 🌾 许可证: MIT 反馈: OpenClaw社区 Discord