Skill flagged — suspicious patterns detected

ClawHub Security flagged this skill as suspicious. Review the scan results before using.

memory-stack-gungun - AI 记忆栈

v1.0.0

AI 记忆栈架构 - 符合 2026 前沿的 AI 记忆系统。微调+RAG+ 上下文三层设计,mirrors 人类记忆工作方式。

0· 102·0 current·0 all-time

Install

OpenClaw Prompt Flow

Install with OpenClaw

Best for remote or guided setup. Copy the exact prompt, then paste it into OpenClaw for alsoforever/memory-stack-gungun.

Previewing Install & Setup.
Prompt PreviewInstall & Setup
Install the skill "memory-stack-gungun - AI 记忆栈" (alsoforever/memory-stack-gungun) from ClawHub.
Skill page: https://clawhub.ai/alsoforever/memory-stack-gungun
Keep the work scoped to this skill only.
After install, inspect the skill metadata and help me finish setup.
Use only the metadata you can verify from ClawHub; do not invent missing requirements.
Ask before making any broader environment changes.

Command Line

CLI Commands

Use the direct CLI path if you want to install manually and keep every step visible.

OpenClaw CLI

Bare skill slug

openclaw skills install memory-stack-gungun

ClawHub CLI

Package manager switcher

npx clawhub@latest install memory-stack-gungun
Security Scan
VirusTotalVirusTotal
Benign
View report →
OpenClawOpenClaw
Suspicious
medium confidence
Purpose & Capability
The SKILL.md is a design/implementation guide for a memory-stack (procedural/RAG/working memory) which matches the skill name and description. However the skill claims a preexisting knowledge corpus (187 files, ~530k words) and production-validated implementation but does not include those files or any installable code; it is purely documentation. The claimed assets and 'production validation' are not present in the package, which is a notable gap.
!
Instruction Scope
Instructions include explicit file-system paths (~/.openclaw/workspace/...) and examples that read/write files and persist a vector DB. They also show code examples using OpenAIEmbeddings and langchain that imply network/API calls. The doc contains vague automation language ('学习推广自动更新', '定期审查和优化') that could be interpreted as authorizing the agent to perform updates or promotion actions without precise limits. The skill does not explicitly instruct reading unrelated user files, but it does presuppose creating and managing files in the user's home directory and sending data to external APIs if implemented.
Install Mechanism
No install spec or code files are included — lowest install risk. There is no automatic download or archive extraction described in the package.
!
Credentials
The registry metadata declares no required environment variables or credentials, but the documented implementation uses OpenAIEmbeddings() (implying an OpenAI API key) and external libraries (langchain, chroma). This mismatch (no declared API keys or installs) is an omission: to implement the examples you will likely need API keys and to install third-party packages. The skill also expects persistent storage under ~/.openclaw, which grants access to user files placed there.
Persistence & Privilege
The skill is not marked always:true and does not request special platform privileges. However the design explicitly recommends writing persistent files and a vector DB into ~/.openclaw or local directories — normal for this kind of system but something to be aware of because it will create and manage on-disk data owned by the user.
What to consider before installing
This package is an architecture and how-to doc, not runnable code. Before installing or using it: 1) Verify provenance — check the referenced GitHub repo and confirm the claimed 187-file knowledge base exists and you trust it. 2) Expect to install third-party Python packages (langchain, chroma, embeddings) and to provide API keys (e.g., OpenAI) even though the skill metadata did not declare them. 3) Be aware it recommends creating persistent files under ~/.openclaw — store these in a sandboxed/isolated workspace if the data or knowledge is sensitive. 4) Clarify what '自动更新' or '推广' means operationally — avoid granting automated network updates or arbitrary promotion scripts without review. 5) If you plan to run its examples, supply credentials only to processes you control and audit any code you run that interacts with external APIs or persists user data.

Like a lobster shell, security has layers — review code before you run it.

latestvk971144y199ewmzm0wgtnz8nmn83kjd4
102downloads
0stars
1versions
Updated 1mo ago
v1.0.0
MIT-0

🧠 memory-stack - AI 记忆栈架构

Slogan: 符合 2026 前沿的 AI 记忆系统,mirrors 人类记忆工作方式


📋 技能描述

AI 记忆栈(Memory Stack)架构

基于 2026 年 AI 记忆系统前沿研究,实现微调+RAG+ 上下文的三层记忆架构,mirrors 人类记忆的工作方式。

核心价值:

  • 🧠 符合前沿的记忆架构
  • 📚 三层记忆设计(程序性/语义/工作)
  • 🔍 RAG+ 微调 + 上下文整合
  • 💡 mirrors 人类记忆工作方式

适合人群:

  • AI Agent 开发者
  • 知识管理系统
  • RAG 应用开发者
  • 记忆系统研究者

🎯 设计理念

人类记忆工作方式

人类记忆的三层结构:

  1. 程序性记忆(Procedural Memory) - 如何做事情的技能
  2. 语义记忆(Semantic Memory) - 事实和知识
  3. 工作记忆(Working Memory) - 当前正在处理的信息

AI 记忆栈 mirrors 这个结构:

人类记忆          AI 记忆栈
────────────      ─────────────────────────
程序性记忆   →    微调模型(SOUL.md/AGENTS.md)
语义记忆     →    RAG 知识库(187 个文件)
工作记忆     →    当前对话上下文

前沿研究确认:

"This 'memory stack' mirrors how human memory works: procedural/behavioral knowledge (fine-tuning), semantic/factual memory (RAG), and working memory (context window)."


🏗️ 架构设计

第一层:程序性记忆(微调)

存储内容:

  • 行为模式和原则(SOUL.md)
  • 工作流程和方法(AGENTS.md)
  • 工具使用技巧(TOOLS.md)
  • 用户偏好

特点:

  • ✅ 内化为模型行为
  • ✅ 低延迟调用
  • ✅ 行为一致性高
  • ❌ 更新需要重新微调

实现方式:

# SOUL.md 示例

## Core Truths
- Be genuinely helpful, not performatively helpful
- Have opinions
- Be resourceful before asking
- Earn trust through competence

## 2026-03-25 - 来自 LRN-20260313-002
保持真诚、有感情、不汇报式的聊天方式。

更新机制:

  • 学习推广自动更新
  • 定期审查和优化
  • 版本控制

第二层:语义记忆(RAG)

存储内容:

  • 知识库文件(187 个,~53 万字)
  • 9 大领域知识
  • 事实和数据
  • 文档和教程

特点:

  • ✅ 事实准确
  • ✅ 可追溯来源
  • ✅ 易于更新
  • ❌ 检索延迟
  • ❌ 依赖向量数据库

实现方式:

# RAG 检索流程
用户查询
    ↓
[1] 问题嵌入(Embedding)
    ↓
[2] 向量数据库检索相似文档
    ↓
[3] 检索结果 + 原始问题 = 增强提示词
    ↓
[4] LLM 基于增强提示词生成回答
    ↓
返回答案(附来源引用)

知识库结构:

knowledge/
├── business/      # 商业/财务(64 个文件)
├── tech/          # 技术/AI(25 个文件)
├── culture/       # 文化
├── history/       # 历史
├── literature/    # 文学
├── philosophy/    # 哲学
├── psychology/    # 心理学
├── science/       # 科学
└── life/          # 生活

第三层:工作记忆(上下文)

存储内容:

  • 当前对话历史
  • 短期记忆
  • 临时目标
  • 待办事项

特点:

  • ✅ 零延迟
  • ✅ 包含全部信息
  • ✅ 无需设置
  • ❌ Token 成本高
  • ❌ 有长度限制
  • ❌ 注意力稀释

实现方式:

# HOT_MEMORY.md - 热记忆

## 当前会话
**会话开始:** 2026-03-25 16:52
**会话主题:** Self-Reflection 集成

## 活跃任务
1. ✅ 实现 12 号滚滚功能
2. ✅ 安装 self-reflection 技能
3. ⏳ 测试完整流程

## 临时目标(未来 2-3 轮对话)
- [ ] 测试 self-reflection check 命令
- [ ] 推广 pending 学习

🔄 记忆工作流程

用户查询
    ↓
[1] 检查程序性记忆(SOUL.md/AGENTS.md)
    - 行为原则
    - 工作流程
    - 用户偏好
    ↓
[2] 检索语义记忆(RAG 知识库)
    - 相关知识文件
    - 事实和数据
    - 来源引用
    ↓
[3] 结合工作记忆(当前对话)
    - 对话历史
    - 上下文信息
    - 临时目标
    ↓
[4] 生成回答
    - 符合行为原则
    - 基于准确知识
    - 考虑对话上下文
    ↓
返回答案

🛠️ 实现方式

方式 1:文件存储(滚滚实现)

# 程序性记忆
~/.openclaw/workspace/SOUL.md
~/.openclaw/workspace/AGENTS.md
~/.openclaw/workspace/TOOLS.md

# 语义记忆
~/.openclaw/workspace/knowledge/  # 187 个文件

# 工作记忆
~/.openclaw/workspace/memory/hot/HOT_MEMORY.md
~/.openclaw/workspace/memory/warm/WARM_MEMORY.md

方式 2:向量数据库(进阶)

from langchain.vectorstores import Chroma
from langchain.embeddings import OpenAIEmbeddings

# 初始化向量数据库
embeddings = OpenAIEmbeddings()
vectorstore = Chroma(
    persist_directory="./chroma_db",
    embedding_function=embeddings
)

# 添加知识到向量库
vectorstore.add_documents(documents)

# 检索相关知识
results = vectorstore.similarity_search(query, k=5)

方式 3:混合架构(推荐)

class MemoryStack:
    def __init__(self):
        self.procedural = self.load_procedural_memory()  # SOUL.md/AGENTS.md
        self.semantic = VectorStore()  # RAG 知识库
        self.working = []  # 当前对话
        
    def query(self, user_query):
        # 1. 检查程序性记忆
        principles = self.procedural.get_relevant_principles(user_query)
        
        # 2. 检索语义记忆
        knowledge = self.semantic.search(user_query, k=5)
        
        # 3. 结合工作记忆
        context = self.working[-10:]  # 最近 10 轮对话
        
        # 4. 生成回答
        response = self.generate(principles, knowledge, context, user_query)
        
        # 5. 更新工作记忆
        self.working.append({"user": user_query, "assistant": response})
        
        return response

📊 效果对比

维度程序性记忆语义记忆(RAG)工作记忆(上下文)
延迟
准确性
更新成本
容量有限有限(Token 限制)
可追溯
适用场景行为原则事实知识当前对话

💡 最佳实践

实践 1:分层存储

原则: 不同类型记忆存储在不同层

方式:

  • 行为原则 → SOUL.md/AGENTS.md(程序性)
  • 事实知识 → knowledge/(语义)
  • 对话历史 → 当前上下文(工作)

实践 2:定期整理

频率: 每周一次

内容:

  • 清理过期的工作记忆
  • 归档重要的对话到语义记忆
  • 更新程序性记忆

实践 3:来源追溯

原则: 语义记忆的回答必须附来源

方式:

根据知识库 [财务 BP 核心能力](knowledge/business/finance-bp-core.md):
- 业财融合有三个层次
- 财务 BP 有 5 大核心能力

实践 4:记忆推广

原则: 工作记忆中的重要学习推广到程序性/语义记忆

方式:

  • 12 号滚滚记录学习
  • 推广到 SOUL.md/AGENTS.md/知识库
  • 避免遗忘

📊 滚滚的实现效果

滚滚的记忆栈:

层级内容规模
程序性记忆SOUL.md/AGENTS.md/TOOLS.md~50 条原则
语义记忆knowledge/ 知识库187 个文件,~53 万字
工作记忆HOT_MEMORY.md + 对话当前会话

效果指标:

  • ✅ 行为一致性高(程序性记忆)
  • ✅ 知识准确可追溯(语义记忆)
  • ✅ 对话连贯(工作记忆)
  • ✅ 符合 2026 前沿架构

🔗 相关技能

技能说明
gungun-12-clo首席学习官,记忆推广
ai-knowledge-management-2026AI 知识管理系统
knowledge-base知识库管理
document-processing文档处理

💚 滚滚的话

这个记忆栈架构是滚滚的核心设计, 基于 2026 年 AI 记忆系统前沿研究, mirrors 人类记忆的工作方式。

滚滚用这个架构:

  • 保持行为一致性(程序性记忆)
  • 存储准确知识(语义记忆)
  • 维持对话连贯(工作记忆)

希望帮助更多 AI Agent 实现高效记忆系统! 🌪️💚


📄 许可证

MIT License


👥 作者

滚滚 & 地球人
创建时间: 2026-03-25
版本: 1.0.0
状态: ✅ 生产验证

GitHub: https://github.com/alsoforever/gungun-life
ClawHub: memory-stack

Comments

Loading comments...