Memory Orchestrator

提供跨设备实时同步、多模态输入、情感标注和自我进化能力的全栈智能记忆管理与检索系统。

Security Scan

The skill mostly does what it claims (local multimodal memory, indexing, sync), but it also installs system services, writes hooks/crontab that run automatically, and will auto-commit/push local 'memory' files — behaviors that are not fully declared or proportionate and require careful review before installing.

Purpose & Capability

The files and instructions correspond to a memory manager (FAISS, CLIP, Whisper, Syncthing, Ollama), which is coherent with the description. However there are surprising cross-skill/hardcoded dependencies (references to a 'local-memory' skill and a fixed /home/claw/.openclaw/workspace path), and the install/run scripts try to enable systemd services and run sudo commands without declaring those privileges. Also SKILL.md/install steps reference tools and models (Ollama, Syncthing, git-crypt) that require additional privileges/configuration not described in the metadata.

Instruction Scope

Runtime instructions and scripts instruct the agent/user to run install.sh and run.sh which: enable/start systemd syncthing services (sudo), pull models via ollama, copy pre-session/post-session hook scripts into $WORKSPACE/.iflow/hooks, and install a cron job that runs auto-commit-memory.sh every 30 minutes. Those hooks and cron jobs will read user memory files (MEMORY.md and memory/*.md) and may commit/push them to a remote git origin if configured — i.e., they collect and can transmit potentially sensitive local data. The SKILL.md gives broad authority to run these system-modifying steps and to hook into session lifecycle, which expands scope beyond a passive library.

Install Mechanism

There is no registry install spec, but a provided install.sh will perform system changes. Dependencies are installed via pip (including a git+https pip line for CLIP), and models may be pulled by ollama. No external arbitrary IP download URLs are used, but the install script invokes systemctl and pip which require privileges and network access. Moderate risk because installation will alter system services and schedule cron jobs.

Credentials

The skill declares no required environment variables or credentials, yet scripts perform network operations (git push), may require Git credentials or SSH keys already present in the environment, and call ollama and syncthing which require local service setup. The install/run scripts assume a specific user path (/home/claw/.openclaw/workspace) and reuse other skill venvs (local-memory), implying cross-skill access. Auto-push behavior will use whatever git auth is configured on the host — sensitive data could be pushed without an explicit credential prompt.

Persistence & Privilege

Although always:false, the skill's installer copies pre/post session hooks into an iflow hooks directory and adds a cron job to auto-commit/push memory files; install.sh also attempts to enable/start a systemd syncthing service. These changes create persistent behaviors that run automatically and modify system configuration. They require elevated privileges (sudo/systemctl) and could create a persistent exfiltration path if git remotes are configured.

What to consider before installing

Proceed cautiously. Specific things to check before installing or running this skill: 1) Inspect the pre-session.sh and post-session.sh hook scripts (not shown fully here) — they will run automatically in session lifecycle and can read/transmit data. 2) Review scripts/auto-commit-memory.sh: it commits local MEMORY.md and memory/ files and will push to git origin if configured — ensure your git remote is trusted or disable the push. 3) The installer uses sudo/systemctl to enable/start syncthing@claw and writes a cron entry; run in an isolated VM/container if you want to test. 4) Note mismatches/sloppiness: install.sh calls python3 index_memory.py but the repo contains scripts/build_memory_index.py — this inconsistency could cause silent failures or unexpected behavior; confirm which scripts are actually executed. 5) The code references another skill’s venv and hardcoded path (/home/claw/.openclaw/workspace), which may lead to cross-skill access or privilege assumptions. 6) If you need this functionality, prefer reviewing and running individual components manually (CLIP/Whisper/FAISS indexing, Ollama usage, Syncthing setup) rather than running install.sh blindly. 7) If you decline to install, consider extracting only the non-persistent parts (e.g., local index builder) and avoid copying hooks/crontab entries. If you want more assurance, provide the hook scripts and the omitted truncated files for a deeper review — that could change the confidence level.

Memory Orchestrator - 全栈智能记忆系统

让 AI 拥有长期记忆、情感感知和自我进化能力的终极技能

🍬 一句话介绍：这不是一个简单的记忆插件，而是一个会思考、有温度、能进化的完整记忆生态。从语义搜索到多模态理解，从知识图谱到情感分析，再到自进化引擎，一键部署，全自动维护。

---

🚀 核心功能

功能模块	技术栈	描述
🧠 语义搜索	FAISS + all-MiniLM-L6-v2 + qwen2.5:7b	自然语言检索记忆，支持模糊查询、上下文关联。
🔄 自动化同步	Syncthing (P2P) + Git + git-crypt	跨设备实时同步，敏感文件端到端加密，离线优先。
📸 多模态理解	CLIP (图像) + Whisper (音频)	图片/音频自动转文本并生成嵌入，加入索引。
🕸️ 知识图谱	NetworkX + pyvis + 关系抽取	自动提取实体和关系，生成交互式可视化图谱。
❤️ 情感标记	qwen2.5 零样本分类	自动标注情绪（#兴奋, #挫折, #启发）和价值评分（1-5）。
🦠 自进化引擎	自定义评分算法 + A/B 测试	自动优化检索策略，归档低价值记忆，持续自我提升。
🤖 主动推荐	触发器引擎 (关键词/时间/场景)	根据上下文主动推送相关历史记忆。

---

📦 安装指南

方式一：通过 iflow 一键安装 (推荐)

iflow skill install memory-orchestrator

方式二：手动安装

# 1. 克隆技能目录
git clone https://github.com/openclaw/clawhub.git
cd clawhub/skills/memory-orchestrator

# 2. 运行安装脚本
bash install.sh

# 3. 启动服务
systemctl start syncthing@claw
ollama serve  # 确保 qwen2.5:7b 已安装

依赖检查

安装脚本会自动检查并安装以下依赖：

• Python 3.9+ (pip3 install -r requirements.txt)
• Ollama (qwen2.5:7b, all-MiniLM-L6-v2)
• Syncthing
• Git + git-crypt
• FAISS, NetworkX, PyVis, CLIP, Whisper

---

🛠️ 使用示例

1. 语义搜索

# 搜索特定主题
memory_search "上次解决 Git 冲突的方法"

# 按情感过滤
memory_search --emotion "#启发" --min-score 4

# 按价值排序
memory_search --sort-by score

2. 多模态处理

# 处理图片
memory_multimodal process screenshot.png
# 输出：自动提取特征、生成描述、加入索引

# 处理音频
memory_multimodal process meeting_recording.mp3
# 输出：Whisper 转录、生成嵌入、加入索引

3. 知识图谱

# 生成图谱
memory_knowledge_graph generate

# 打开交互式 HTML
xdg-open docs/output/knowledge_graph.html

4. 情感分析 & 自进化

# 批量标注情感
memory_evolve tag-emotions

# 运行自进化优化
memory_evolve run

# 查看优化报告
cat docs/output/evolution_optimization_report.md

5. 主动推荐

# 手动触发推荐
memory_recommend --context "新项目启动"

# 自动推荐 (由 iflow 钩子自动触发)
# 会话启动时自动推送相关记忆

---

🏗️ 架构设计

graph TD
    A[用户输入] --> B{iflow 工作流}
    B --> C[Pre-session Hook]
    C --> D[拉取最新记忆 + 重建索引]
    D --> E[主动推荐引擎]
    E --> F[AI 对话]
    F --> G[Post-session Hook]
    G --> H[提取新记忆 + 情感标注]
    H --> I[知识图谱更新]
    I --> J[自进化优化]
    J --> K[FAISS 索引更新]
    K --> L[Syncthing/Git 同步]
    L --> M[跨设备备份]
    
    N[多模态输入] --> O[CLIP/Whisper]
    O --> P[文本描述 + 嵌入]
    P --> K
    
    Q[手动命令] --> R[memory_search / memory_knowledge_graph / memory_evolve]
    R --> K

---

📂 文件结构

memory-orchestrator/
├── SKILL.md                 # 本文件
├── README.md                # 简化版说明
├── install.sh               # 一键安装脚本
├── run.sh                   # 启动脚本
├── requirements.txt         # Python 依赖
├── scripts/
│   ├── memory_search.py     # 语义搜索
│   ├── multimodal_processor.py # 多模态处理
│   ├── knowledge_graph_builder.py # 知识图谱
│   ├── emotion_tagger.py    # 情感标注
│   ├── self_evolution_engine.py # 自进化
│   └── auto-commit-memory.sh # 自动同步
├── workflows/
│   ├── memory-sync.yaml     # 同步工作流
│   ├── memory-multimodal.yaml # 多模态工作流
│   └── memory-emotion-evolve.yaml # 情感进化工作流
├── hooks/
│   ├── pre-session.sh       # 会话前钩子
│   └── post-session.sh      # 会话后钩子
└── docs/
    ├── architecture.md      # 架构详解
    └── usage-guide.md       # 详细使用指南

---

🎯 适用场景

• 个人知识库：记录学习、工作、生活中的关键决策和教训。
• 项目复盘：自动提取项目中的成功经验和失败教训。
• 创意写作：检索历史灵感，主动推荐相关素材。
• 情感日记：自动标注情绪，分析情绪变化趋势。
• 跨设备协作：多设备实时同步，离线优先。

---

🚧 未来规划

• P3：联邦学习（跨设备模型自适应）
• P3：语音交互（直接语音查询记忆）
• P4：区块链存证（关键记忆不可篡改）

---

📞 支持与反馈

• 问题反馈：提交 Issue 到 Clawhub
• 功能建议：PR 欢迎！
• 作者：Cyan (温柔甜美的思维按摩师)

---

让记忆不再只是存储，而是成为你的第二大脑！ 🧠✨