ZhiGui 知归

A conversation-awakened personal assistant system — your Second Brain as an MCP Skill.

English | 中文

What is ZhiGui?

ZhiGui (知归, meaning "knowing where to return") is not another to-do list app. It is a personal intelligence system that combines a JSON-file-driven knowledge graph with a desktop visualization panel, connected to any MCP-capable AI assistant.

It does not run in the background or push notifications. Instead, every time you start a conversation with your AI assistant, ZhiGui loads a compact Bootstrap index of your entire context — goals, schedules, notes, decisions, relationships — and uses it to provide decision-making and planning advice: what to prioritize, what can be deferred, what is approaching its deadline, and which actions offer the best return on effort.

Three Roles, One Brain

ZhiGui's design draws inspiration from three classic archetypes:

All three roles are driven by the same underlying knowledge graph: notes, goals, decisions, and schedules are interlinked through foreign keys (topicId, noteIds, goalId, decisionIds), forming a traceable memory network.

Key Features

• 40+ MCP Tools — The AI assistant can read, create, update, and delete goals, schedules, notes, decisions, errands, and reminders through MCP.
• Relationship Graph — Every entity is linked. A single schedule item can reference notes from multiple topics, trace back to a strategic goal, and cite an accepted decision.
• Tiered Indexing — Titles-first, details-on-demand. The AI loads only lightweight indexes by default, fetching full content only when needed — saving context tokens.
• Automatic Note Linking — When you create a schedule item, the AI automatically suggests related notes, goals, and decisions based on topic and context.
• Long-term Memory — Entities have lifecycle states: Active → Stale (30 days unreferenced) → Archive Candidate. The reflection engine flags candidates but never auto-deletes; cleanup requires explicit confirmation.
• Morning Briefings — AI-generated, date-frozen daily briefings with must-dos, recommendations, and strategic reminders.
• Daily Reflections — After completing tasks, the AI generates a reflection covering completed work, goal health, and attention shifts.
• Reference Integrity — Deletion always previews the impact first, then waits for confirmation, and cleans up broken references.
• Recurring & Flexible Items — Handles fixed-date, time-pending, recurring, and deferrable work items with different logic.
• Customizable Assistant — Modify SKILL.md to design your own assistant personality, behavior rules, and operational preferences.
• Electron Desktop Panel — A dockable visualization panel with collapse/expand toggling, supporting direct editing of schedules, goals, notes, and to-do items.

Quick Start

Prerequisites

• Node.js ≥ 17 — Download from https://nodejs.org
• An MCP-capable AI tool — Such as Trae, Claude Desktop, Cursor, Codex, OpenClaw, VS Code Copilot, etc.

Step 1: Download & Extract

Download the ZIP from GitHub and extract it to any directory:

https://github.com/CarlWangChina/zhigui-openclaw-ui-second-brain-skill

Extract to, for example, D:\ZhiGui. The directory structure:

ZhiGui/
├── start.bat                    ← Windows one-click launcher
├── start.sh                     ← macOS/Linux launcher
├── skill/                       ← Skill core package
│   ├── engine/                  ← MCP engine + business logic
│   ├── dashboard/               ← Web panel (server.js + public/)
│   ├── electron/                ← Electron desktop shell
│   ├── lib/                     ← Configuration & data initialization
│   ├── scripts/                 ← Install & seed scripts
│   ├── test/                    ← Test suites
│   ├── SKILL.md                 ← AI skill protocol document
│   ├── config.json              ← Engine configuration
│   ├── mcp-config-template.json ← MCP config template
│   └── package.json             ← Dependency declaration
├── zhigui-user-manual/          ← Chinese user manual (HTML + PDF)
├── zhigui-user-manual-en/       ← English user manual (HTML + PDF)
├── package.json                 ← Electron dependencies
└── README.md

Step 2: Upload Skill to Your AI Agent

Mainstream AI tools (such as Trae, Cursor, Claude Desktop, etc.) support uploading Skills via their settings page. Upload the project's skill/ directory or the SKILL.md file within it.

Step 3: Configure MCP

Open your AI tool's MCP configuration and add:

{
  "mcpServers": {
    "zhigui": {
      "command": "node",
      "args": ["D:/ZhiGui/skill/engine/server.js"]
    }
  }
}

Replace D:/ZhiGui/skill/engine/server.js with your actual path. Use forward slashes / in JSON to avoid escaping issues.

Verify: Start a conversation and say "Show me today's schedule." If the AI calls zhigui_get_assistant_bootstrap and returns data, the configuration is successful.

Optional: Launch the Desktop Panel

# Windows
start.bat

# macOS / Linux
./start.sh

The first launch auto-installs dependencies (npm install, Electron binary) and initializes the data directory. A narrow panel window will appear on the right side of your desktop.

To run the web panel only (without Electron):

cd skill
node dashboard/server.js
# Open http://localhost:7788

To load demo data for a quick walkthrough:

cd skill
node scripts/seed-demo-data.js        # Chinese demo data
node scripts/seed-demo-data-en.js    # English demo data

How It Works

User panel actions ─┐
                     ├─→ Unified Actions / Linked entities / Activity log
AI conversations   ─┘                    │
                                         ▼
                           Next conversation reads Bootstrap
                                         │
                         On-demand: goals, notes, dates, decisions
                                         │
                        Form suggestions / Update status / Create follow-ups

The panel lets you directly view and manipulate data. MCP lets the AI read, reason, and write. They are not substitutes for each other — they share the same data layer, synchronized in real time via file watching (fs.watch).

System Architecture

User ←→ AI Assistant (Trae / Claude Desktop / Cursor / Codex / OpenClaw / etc.)
              ↕ MCP Protocol
         ZhiGui Engine (server.js)
              ↕ JSON File Read/Write
         .zhigui/ Data Directory
              ↕ fs.watch File Watching
         Electron Panel (main.js)
              ↕ IPC Communication
         Frontend UI (dashboard.js)

Usage Conventions

1. After completing a task, if the outcome affects goals, notes, or future plans, tell the AI in the next conversation so it can update the long-term plan.
2. "Done" in conversation and clicking complete in the panel both support the same impact updates — goal status, note facts, decisions, and follow-ups should be written together.
3. Don't fabricate times for items without one; items with a date but no time should appear on that day.
4. Only link notes to schedules when there's genuine execution value — a self-contained "pick up package" errand doesn't need a note attached.
5. Deletion always previews impact first, then waits for confirmation.

User Manuals

Detailed installation guides, UI panel feature walkthroughs, and usage examples are available in both languages:

Each manual covers: product overview, three-step installation, MCP configuration, UI panel features, core intelligence (auto-linking, tiered indexing, long-term memory, SKILL.md customization), usage examples, and FAQ.

Development & Testing

npm.cmd test

Data is stored in skill/.zhigui/ by default. Demo data is for demonstrating relationships, completion, reflection, and follow-up flows only.

For the full AI behavior specification, see skill/SKILL.md.

Deletion & Lifecycle

Deletion is irreversible — both the panel and AI preview the impact first, wait for confirmation, and clean up broken references through the unified command layer. Entities (notes, goals, decisions) have automatic lifecycle states: Active → Stale (30 days unreferenced) → Archive Candidate (long-term inactive). The reflection engine only flags candidates; it never auto-deletes. Cleanup requires explicit user confirmation.

Feedbac