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 generally matches its 'continuous learning' purpose, but there are several implementation and scope issues you should review (file references missing, broad file reads, periodic network calls) before installing or running it with real data or credentials.

Purpose & Capability

Name/description align with code: the package reads local memory/docs, writes learning files, syncs to a MemPalace skill, and optionally calls a BigModel API for analysis. However install scripts reference helper files (schedule_start.bat, install_autostart_en.bat) that are not present in the file manifest — this is an inconsistency that will break the advertised 'one-click' autostart flow. Overall capabilities requested (reading/writing local docs, calling MemPalace and MiniMax, sending notifications) are plausible for the stated purpose.

Instruction Scope

Runtime logic and SKILL.md instruct the agent to automatically read many local files (SOUL.md, AGENTS.md, MEMORY.md, TOOLS.md, .learnings/*, memory/YYYY-MM-DD.md) and explicitly allow reading skills/<skill>/SKILL.md. That is within a 'learning' feature but grants broad local file access (including other skills) which can expose sensitive info. The scheduled tasks automatically call an external API (MiniMax) and use 'openclaw message send' to push notifications. The instructions also instruct automatic modification of core documents — ensure you trust the code and contents that will be aggregated.

Install Mechanism

There is no remote download/install spec (instruction-only + local Python scripts + requirements.txt). That lowers third-party code risk, but the package expects Python dependencies (chromadb, requests, openai). No external URLs or archive extraction are used. Note: install_cron.py references batch scripts that are absent from the manifest; the 'pure Python scheduler' approach runs as a long-lived process rather than integrating with system cron/task scheduler.

Credentials

The skill declares no required env vars, but it expects an API key placed in config/dream_config.json (analysis.api_key) and uses an 'openclaw' CLI to send notifications (which depends on the host's OpenClaw credentials/config). It also reads other skills' SKILL.md files and workspace documents. Storing API keys in plaintext config and allowing automatic reads/writes to workspace files is functional but increases exposure; the scope of file access is broader than a minimal notifier and could leak sensitive content to external services if misconfigured.

Persistence & Privilege

The skill does not force 'always: true' and does not request special registry privileges. It installs a long-running Python scheduler (schedule_runner.py) intended to run continuously or at boot (via missing autostart scripts). Autonomous invocation is allowed (disable-model-invocation is false), which combined with periodic jobs and broad file I/O increases impact — this is expected for a continuous-learning skill but worth caution.

Scan Findings in Context

[subprocess.run_usage] expected: Used to run 'openclaw message send' and to start Python scripts for scheduled tasks. Expected for notification and scheduling but verify openclaw CLI privileges and behavior.

[requests_http_call] expected: Script calls external MiniMax API via requests.post for analysis. Expected given the feature, but any data sent to that API may include user conversation content—verify the destination and trustworthiness of the API key.

[workspace_file_reads_writes] expected: The code reads/writes many workspace documents (.learnings/, memory/, SOUL.md, AGENTS.md, other skills' SKILL.md). This is functional for the skill but increases the risk of exposing sensitive workspace contents.

[sys_path_insertion_import_other_skill] expected: The dream and sync scripts insert a skills/mempalace path and import mempalace.mcp_server.handle_request. This is necessary for integration but you should audit the mempalace skill and ensure it is trusted.

What to consider before installing

This skill is mostly coherent with 'continuous learning' behavior but exercises broad local file access and periodic network calls. Before installing or running it: 1) Inspect and audit the mempalace skill it requires; do not install this unless mempalace is trusted. 2) Do not put sensitive API keys or secrets into config/dream_config.json unless you trust the target MiniMax API; prefer environment-based secret management. 3) Review which workspace files (SOUL.md, AGENTS.md, other skills' SKILL.md) exist and whether they contain secrets — the skill will read and may copy their contents to external services. 4) Verify the openclaw CLI configuration and who receives notifications (config.notification.to). 5) Note two referenced helper scripts (schedule_start.bat, install_autostart_en.bat) are missing — the autostart flow may be incomplete; run the scheduler in an isolated test workspace first. 6) If you plan to run periodic background tasks, consider running them in an isolated environment or container, and monitor outbound network traffic to the configured API endpoints. If you want, provide the mempalace skill source and the actual target MiniMax endpoint to increase my confidence.

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（Python调度器 v2.0）

所有平台（推荐使用Python调度器）:

python setup/install_cron.py

Python调度器优势（v2.0新特性）：

• ✅ 纯Python实现：无需系统权限，不依赖任务计划程序/cron
• ✅ 开机自动启动：支持Windows自动启动配置
• ✅ 微信通知完善：任务执行后自动发送通知
• ✅ 精确时间匹配：每小时轮询，准确到分钟
• ✅ 跨平台一致：Windows/Linux/Mac体验一致

手动启动：

# Windows
启动脚本: schedule_start.bat

# Linux/Mac
python setup/schedule_runner.py

设置开机自动启动：

# Windows
双击: install_autostart_en.bat

# Linux/Mac (TODO: 添加systemd服务支持)

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