AI Memory Manager

English documentation is below the Chinese section.

"给你的 LLM 装上一个永远不会失忆、不写烂代码的本地大脑。"

AI Memory Manager 是一款 VSCode 扩展，通过对话式 AI 提取架构规则并"幽灵注入"到 .cursorrules，让 Cursor、Continue 等 AI 代码助手始终遵守你制定的工程规范。100% 本地存储，支持自带 API Key（BYOK）。

✨ 核心功能 (v1.1.6)

🔧 基础功能

• BYOK 多 Provider 支持：OpenAI / Anthropic / Google Gemini / Ollama / DeepSeek / 通义千问 / 豆包 / 自定义端点，一键切换
• AI 规则蒸馏：口语化描述 → LLM 提炼 → 专业架构规则，无需精准措辞
• 幽灵注入：规则自动写入 .cursorrules，AI 助手即时生效，零感知

🛡️ 商业级防御

• 机密扫描器：写入本地前强制扫描，sk-xxx / password= / MongoDB URI 自动脱敏，防止密钥泄漏
• 外部变更监控：chokidar 监听 .prisma / schema.sql，数据库被外部工具修改时立即告警
• 熔断限流器：单次对话 AI 工具调用上限 5 次，防止死循环耗尽 API 额度
• 决策日志：所有 AI 行为写入 decision_log.md，支持历史溯源 Debug

🔍 智能检索

• 语义模糊搜索（Fuse.js）：跨 entities/ 目录模糊匹配，召回最相关知识
• 向量语义检索：基于 WebAssembly ONNX 运行时，跨平台本地向量嵌入
• Function Calling：LLM 可主动调用 search_memory 工具检索本地记忆库
• 双模式分流：自动区分"规则提取"和"知识查询"两种意图

📦 配置管理

• 一键导出：整个 .aimemory/ 目录打包为 .zip，便于迁移备份
• 一键导入：还原 zip 包，UI 自动刷新，跨机器同步配置
• 全局持久化：API Key / Model / Base URL 跨 session 保持

🌍 跨平台支持

• Windows / macOS / Linux 完整支持：向量检索引擎基于 WebAssembly，无平台原生二进制依赖

📁 数据存储结构

your-workspace/
├── .aimemory/
│   ├── config/
│   │   └── global_rules.json    ← AI 提取的架构规则库
│   ├── entities/                ← 结构化数据字典（供语义搜索）
│   ├── apis/                    ← 接口契约与签名
│   └── contexts/
│       └── decision_log.md      ← 完整 AI 决策历史日志
└── .cursorrules                 ← 自动同步，Cursor/Continue 读取

🚀 快速开始

1. 点击 Activity Bar 中的 AI Memory Manager 图标打开侧边栏
2. 在 Provider Presets 选择你的服务商（如 OpenAI）
3. 填入 Base URL、Model Name 和 API Key，点击 Sync Configuration
4. 在 Ghost Workspace 输入框描述你的规则：
   • 提取规则："以后所有接口返回值必须用 Result<T, E> 包裹"
   • 查询记忆："我们的 User 实体有哪些字段？"
5. 点击 🤖 Distill / Search，规则自动注入并显示在 Active Brain Atlas

🏗️ 版本历史

详细进度记录请见 PROGRESS.md

📦 打包发布

# 安装打包工具
npm install -g @vscode/vsce

# 编译
npm run compile

# 打包为 .vsix
npx @vscode/vsce package

# 发布到 Marketplace
npx @vscode/vsce publish

AI Memory Manager (English)

"Equip your LLM with a local brain that never forgets and never writes bad code."

AI Memory Manager is a VSCode extension that uses conversational AI to extract architectural rules and "ghost-inject" them into .cursorrules. This ensures AI coding assistants like Cursor or Continue consistently follow your specific project standards. 100% local storage, supports Bring Your Own Key (BYOK).

✨ Core Features (v1.1.2)

🔧 Basic Features

• BYOK Multi-Provider Support: OpenAI / Anthropic / Google Gemini / Ollama / DeepSeek / Qwen / Doubao / Custom endpoints with 1-click switching.
• AI Rule Distillation: Natural language → LLM extraction → Professional architecture rules (no prompt engineering required).
• Ghost Injection: Rules are auto-written to .cursorrules. AI assistants instantly adopt them seamlessly.

🛡️ Enterprise Defense

• Secret Scanner: Mandatory scanning before local writes. sk-xxx/password=/MongoDB URLs are redacted to prevent credential leaks.
• External Change Monitoring: chokidar listens to .prisma/schema.sql. Immediate warnings if database schemas are modified externally.
• Circuit Breaker: AI tool calling is limited to 5 loops per session to prevent burning through your API quotas.
• Decision Logger: All AI reasoning is safely logged into decision_log.md for historical debugging and trace-backs.

🔍 Smart Search

• Fuzzy Semantic Search (Fuse.js): Scopes across the entities/ directory to recall the most contextually relevant knowledge.
• Vector Semantic Search: Local embedding via WebAssembly ONNX runtime — cross-platform, no native binaries required.
• Function Calling: LLM can natively invoke search_memory to fetch from local storage.
• Dual Intent Routing: Intelligently differentiates between "Rule Extraction" and "Knowledge Query" intent.

📦 Config Management

• Export in 1-click: Pack the entire .aimemory/ payload into a .zip for team synchronization or backups.
• Import in 1-click: Rehydrate from a zip package with instant UI refreshing.
• Global Persistence: Persist your API Keys/Models across VSCode sessions.

🌍 Cross-Platform

• Windows / macOS / Linux: Vector search engine runs on WebAssembly — no platform-specific native binary dependencies.

📁 Storage Architecture

your-workspace/
├── .aimemory/
│   ├── config/
│   │   └── global_rules.json    ← AI extracted rules
│   ├── entities/                ← Structured Data Dictionary
│   ├── apis/                    ← API Signatures / Contracts
│   └── contexts/
│       └── decision_log.md      ← Action traces
└── .cursorrules                 ← Auto-synced payload context

🚀 Quick Start

1. Click the AI Memory Manager icon in the Activity Bar to open the sidebar.
2. Under Provider Presets, select your AI model provider (e.g. OpenAI).
3. Fill in the Base URL, Model Name, and API Key, then click Sync Configuration.
4. Describe your rule in the Ghost Workspace:
   • Extraction: "All API responses must be wrapped in Result<T, E>"
   • Querying: "What fields does our User entity have?"
5. Click 🤖 Distill / Search. The semantic rule instantly shows up on your Active Brain Atlas.

Made with ❤️ for AI Engineers and Architects. 100% Local Storage · Zero Data Collection · BYOK Pattern

Enjoying AI Memory Manager? ☕ Buy me a coffee to support further development!