Empirehacks 2026
Mnemosyne β Agent-Agnostic Knowledge Management Layer based on DIKW Model. EmpireHacks 2026 Track 3.
Ask AI about Empirehacks 2026
Powered by Claude Β· Grounded in docs
I know everything about Empirehacks 2026. Ask me about installation, configuration, usage, or troubleshooting.
0/500
Reviews
Documentation
ποΈ Mnemosyne
Persistent Memory for AI Agents
Your AI tools generate valuable knowledge every session β then forget everything. Mnemosyne remembers.
Getting Started Β· Architecture Β· How It Works Β· API Reference Β· Contributing
π― The Problem
Every time you close an AI session, all context is lost:
- Claude Code doesn't remember what you researched yesterday
- Codex can't access insights from your Claude sessions
- Valuable decisions, patterns, and lessons are buried in chat logs nobody revisits
- You repeat yourself across sessions, re-explaining context every time
The more you use AI tools, the more knowledge you lose.
π‘ The Solution
Mnemosyne is an agent-agnostic knowledge management layer that sits between your AI tools and creates persistent, cross-platform memory.
It automatically:
- Captures raw conversation data from Claude Code and Codex via file watchers & hooks
- Transforms it through a DIKW pipeline (Data β Information β Knowledge β Wisdom)
- Visualizes everything as an interactive knowledge graph
- Feeds it back into your AI tools via MCP (Model Context Protocol)
The more you use AI tools, the smarter they become.
β¨ Key Features
π Auto-Capture Engine
File watchers monitor ~/.claude/projects/ and ~/.codex/sessions/ in real-time. When a session completes, Mnemosyne automatically ingests the conversation β no manual export needed.
π§ DIKW Transform Pipeline
Raw AI conversations are transformed through four levels using LLM-powered reflection:
βββββββββββ βββββββββββββββ βββββββββββββ ββββββββββ
β Data β βββΆ β Information β βββΆ β Knowledge β βββΆ β Wisdom β
β (raw) β β (context) β β (patterns)β β (meta) β
βββββββββββ βββββββββββββββ βββββββββββββ ββββββββββ
Tool calls, What happened, Reusable When & why
file excerpts why it matters techniques to apply them
π Cross-Platform Knowledge Graph
Interactive D3.js force-directed graph with color-coded DIKW nodes. Pan, zoom, click to explore connections across all your AI sessions.
π MCP Feedback Loop
Mnemosyne exposes an MCP server that feeds accumulated knowledge back into Claude Desktop, Claude Code, and any MCP-compatible client. Your AI assistant gains persistent memory.
π οΈ Platform Hub
Sync skills, MCP servers, and configurations between Claude Code and Codex with one click. Right-click context menus for quick actions.
ποΈ Architecture
flowchart TB
subgraph AI_Platforms["AI Platforms"]
CC["Claude Code"]
CX["Codex"]
CD["Claude Desktop"]
end
subgraph Capture["Auto-Capture Layer"]
FW["File Watchers\n~/.claude/ ~/.codex/"]
HK["Session Hooks\npost-tool-use, stop"]
end
subgraph Core["Mnemosyne Core"]
SRV["Express + WebSocket\nServer"]
DIKW["DIKW Transform\nEngine"]
DB[("SQLite\nDatabase")]
REC["Recommendation\nEngine"]
end
subgraph Interface["User Interface"]
KG["Knowledge Graph\nD3.js Visualization"]
PH["Platform Hub\nSync & Manage"]
INB["Inbox\nNew Sessions"]
end
subgraph Feedback["MCP Feedback Loop"]
MCP["MCP Server\n(stdio)"]
end
CC -->|session files| FW
CX -->|session files| FW
CC -->|hooks| HK
FW --> SRV
HK --> SRV
SRV --> DIKW
DIKW --> DB
DB --> REC
SRV --> KG
SRV --> PH
SRV --> INB
DB --> MCP
MCP -->|knowledge context| CD
MCP -->|knowledge context| CC
style AI_Platforms fill:#1a1a2e,stroke:#667eea,color:#fff
style Capture fill:#16213e,stroke:#f39c12,color:#fff
style Core fill:#0f3460,stroke:#e94560,color:#fff
style Interface fill:#1a1a2e,stroke:#00d2ff,color:#fff
style Feedback fill:#16213e,stroke:#a29bfe,color:#fff
π¬ How It Works
1. Session Capture
Mnemosyne monitors AI tool directories using chokidar file watchers:
// Watches for new/modified session files
watch('~/.claude/projects/**/*.jsonl') // Claude Code sessions
watch('~/.codex/sessions/**/*.jsonl') // Codex sessions
When a session file changes, the watcher automatically parses tool calls, responses, and metadata into Data nodes.
2. DIKW Transformation
The engine processes Data nodes through three LLM-powered reflection stages:
| Stage | Input | Output | Example |
|---|---|---|---|
| D β I | Raw tool calls & responses | Contextual summary | "User researched WebAssembly edge computing trends" |
| I β K | Information nodes | Reusable patterns | "When evaluating new tech, compare: community size, benchmark data, production adoption" |
| K β W | Knowledge nodes | Strategic judgment | "Prioritize technologies with both strong benchmarks AND growing community β performance alone isn't enough" |
3. Knowledge Graph
All nodes and connections are visualized in a real-time D3.js force-directed graph:
- π£ Data β Raw session artifacts
- π΅ Information β Contextualized insights
- π’ Knowledge β Reusable patterns & techniques
- π‘ Wisdom β Meta-level judgment & strategy
4. MCP Feedback
The MCP server exposes three resources to any connected AI client:
mnemosyne://knowledge-nodes β All Knowledge-level insights
mnemosyne://wisdom-nodes β All Wisdom-level insights
mnemosyne://graph-summary β Full graph overview
When Claude Desktop connects, it can access your accumulated knowledge as live context.
π Quick Start
Prerequisites
- Node.js 18+ and npm
- Git
- (Optional) Claude Code and/or Codex installed for auto-capture
Installation
# Clone the repository
git clone https://github.com/ChangXiang-SCU/empirehacks-2026.git
cd empirehacks-2026
# Install dependencies
npm install
cd backend && npm install && cd ..
cd frontend && npm install && cd ..
Run
# Start the backend (port 3001)
npm run backend
# In another terminal, start the frontend (port 5173)
npm run frontend
Open http://localhost:5173 β you'll see the Mnemosyne knowledge graph.
Production Build
npm run build:frontend
npm start
# Frontend is served from backend at http://localhost:3001
Connect to Claude Desktop (MCP)
Add to your Claude Desktop config (claude_desktop_config.json):
{
"mcpServers": {
"mnemosyne": {
"command": "node",
"args": ["<path-to-repo>/backend/mcp-stdio-server.mjs"]
}
}
}
Restart Claude Desktop β you'll see "Add from mnemosyne" in the + menu.
π Project Structure
mnemosyne/
βββ backend/
β βββ server.js # Express + WebSocket + DIKW engine + API
β βββ mcp-stdio-server.mjs # MCP server for Claude Desktop integration
β βββ lib/ # Modular utilities
β βββ db/ # SQLite database (auto-created)
βββ frontend/
β βββ src/
β β βββ components/
β β β βββ MindPalace.jsx # D3 force-directed knowledge graph
β β β βββ ConnectPanel.jsx # Platform Hub (skills, MCPs sync)
β β β βββ NodeCard.jsx # DIKW node detail view
β β β βββ Sidebar.jsx # Project filter panel
β β β βββ InboxPanel.jsx # Unclassified session inbox
β β βββ App.jsx
β βββ dist/ # Production build output
β βββ vite.config.js
βββ hooks/ # Claude Code hook scripts
β βββ post-tool-use.sh # Captures each tool call
β βββ stop.sh # Triggers DIKW reflection on session end
β βββ session-start.sh # Loads context from knowledge graph
βββ demo/ # Demo seed data
βββ launcher.mjs # Auto-launcher script
βββ package.json
βββ LICENSE
βββ CONTRIBUTING.md
π‘ API Reference
REST Endpoints
| Method | Endpoint | Description |
|---|---|---|
GET | /api/graph | Fetch full DIKW graph (nodes + connections) |
GET | /api/nodes?type=K | Filter nodes by DIKW type |
POST | /api/import | Upload ChatGPT/Claude export (multipart) |
POST | /api/dikw/transform | Trigger DIKW pipeline on Data nodes |
POST | /api/hook/tool-use | Claude Code hook endpoint |
GET | /api/platform/skills | List skills from Claude Code & Codex |
GET | /api/platform/mcps | List MCP servers from both platforms |
POST | /api/platform/mcps/sync | Sync MCP server between platforms |
WebSocket Events
| Event | Direction | Description |
|---|---|---|
node:added | Server β Client | New DIKW node created |
connection:added | Server β Client | New edge in graph |
dikw:transform:complete | Server β Client | Pipeline finished |
filewatcher:new-session | Server β Client | New session detected |
MCP Resources
| URI | Description |
|---|---|
mnemosyne://knowledge-nodes | All Knowledge-level nodes |
mnemosyne://wisdom-nodes | All Wisdom-level nodes |
mnemosyne://graph-summary | Aggregated graph overview |
π€οΈ Roadmap
- Auto-capture from Claude Code & Codex
- DIKW transform pipeline with LLM reflection
- Interactive knowledge graph (D3.js)
- MCP server for Claude Desktop integration
- Platform Hub with cross-platform sync
- Claude Code hooks (post-tool-use, stop, session-start)
- ChatGPT export import support
- Multi-user collaboration
- Custom DIKW transformation rules
- Knowledge graph search & semantic query
- Mobile-friendly responsive UI
- Plugin system for custom data sources
π€ Contributing
Contributions are welcome! See CONTRIBUTING.md for guidelines.
π License
MIT β free for personal and commercial use.
Built with π§ for EmpireHacks 2026 β Track 3: The Sidekick
"The more you use AI tools, the smarter they become."
