🧠 Vibe Memory

Universal Long-Term Memory & Codebase AST Intelligence Protocol for AI Coding Agents

Quit Claude Code mid-task, switch to Antigravity or Cursor in the same directory, cut LLM tokens by 97%, and continue coding without re-explaining the architecture, bug discoveries, or user preferences.

🌐 Live Interactive Memory & AST Visualizer

Test the visual memory lineage graph, search indexed codebase symbols, and generate live handoff snapshots:
👉 https://shahrukh-hack.github.io/vibe-memory/

Related MCP server: Paparats MCP

⚡ The Core Problem: Why AI Agents Fail Without Memory

When building production software with AI coding agents (Antigravity, Cursor, Claude Code, OpenAI Codex, Aider), developers face three major bottlenecks:

1. 🔴 The "Fresh Chat Amnesia" Problem: Every time a context window resets or you open a new conversation, the AI forgets your architectural decisions, design tokens, and the edge-case bugs you solved yesterday.

2. 💸 The "Raw File Dumping" Token Tax: When an agent needs to understand a component or function, it reads dozens of full source code files—burning 15,000+ to 50,000+ context tokens and slowing down responses.

3. 🔒 Tool Lock-In & Vendor Fragmentation: Switching from Antigravity (planning) to Cursor (editing) or Claude Code (CLI) requires manually re-prompting and explaining the entire task from scratch.

🚀 The Dual-Engine Solution

Vibe Memory unifies two world-class paradigms into a single, vendor-neutral protocol:

                          ┌────────────────────────────────────────────────────────┐
                          │                VIBE MEMORY DUAL ENGINE                 │
                          └───────────────────────────┬────────────────────────────┘
                                                      │
                       ┌──────────────────────────────┴──────────────────────────────┐
                       ▼                                                             ▼
┌──────────────────────────────────────────────┐              ┌──────────────────────────────────────────────┐
│  ENGINE 1: CONTEXT & HANDOFF MEMORY          │              │  ENGINE 2: AST & SYMBOL KNOWLEDGE GRAPH      │
│  • Architectural Decision Records (ADR)      │              │  • Component & Function Hierarchies          │
│  • Bug fix pattern discoveries               │              │  • Call graphs & export dependency maps      │
│  • Anti-AI slop user preferences             │              │  • 97% Context Token Reduction               │
│  • Seamless cross-agent handoff snapshots    │              │  • Sub-millisecond symbol queries            │
└──────────────────────────────────────────────┘              └──────────────────────────────────────────────┘

🏗️ Architecture & Cross-Tool Flow

graph LR
    A[Antigravity Agent] -->|1. Commit Architectural Decisions| M[AGENT_MEMORY.md Store]
    B[Claude Code CLI] -->|2. Commit Bug Fix Patterns| M
    C[Cursor IDE] -->|3. Read User Intent & Tokens| M
    M -->|4. MCP stdio JSON-RPC| D[Vibe Memory MCP Server]
    D -->|5. Instant AST & Symbol Lookup| E[Any Coding Agent]
    M -->|6. Live Lineage Graph| F[Web Visualizer UI]

📦 How to Use in Ongoing / Existing Projects

Step 1: Initialize in 10 Seconds

Run this single command in the root of any new or existing repository:

npx vibe-memory init

This creates the standard AGENT_MEMORY.md file formatted with 4 persistent sections:

• 📐 1. Architectural Decisions (ADR): Framework choices, state management patterns, and database schemas.

• 🐛 2. Bug Discoveries & Learnings: Edge-case API quirks, platform fixes, and dependency constraints.

• 👤 3. User Preferences: Anti-AI slop design tokens, styling rules, and coding conventions.

• 🔄 4. Active Handoff Checkpoint: Current milestone status, completed steps, and immediate next actions.

Step 2: Connect via Model Context Protocol (MCP)

Add vibe-memory to your AI coding tools to enable automated memory recall and AST queries:

🔹 Antigravity (~/.gemini/antigravity/mcp-config.json):

{
  "mcpServers": {
    "vibe-memory": {
      "command": "npx",
      "args": ["-y", "vibe-memory", "mcp"]
    }
  }
}

🔹 Cursor (.cursor/mcp.json):

{
  "mcpServers": {
    "vibe-memory": {
      "command": "npx",
      "args": ["-y", "vibe-memory", "mcp"]
    }
  }
}

🔹 Claude Code (~/.claude.json):

{
  "mcpServers": {
    "vibe-memory": {
      "command": "npx",
      "args": ["-y", "vibe-memory", "mcp"]
    }
  }
}

🔄 The Cross-Agent Handoff Workflow

When you want to switch from Antigravity to Cursor or Claude Code mid-task:

1. Tell your active agent: "Create handoff snapshot for Cursor."

2. The agent appends a structured checkpoint to AGENT_MEMORY.md:

<!-- AGENT_HANDOFF_SNAPSHOT: ANTIGRAVITY ➔ CURSOR -->
# 🔄 Context Handoff (2026-08-16T21:45:00Z)

## 🎯 Active Objective
Finalizing the real-time fleet telemetry map and vehicle status filter.

## ✅ Accomplished So Far
- Configured Leaflet container and Tailwind design tokens
- Built GPS ingestion pipeline in TypeScript
- Verified Adelaide CBD and Barossa Valley route corridors

## 🚀 Immediate Next Actions for Cursor
- Connect the live incident marker overlay in src/components/CorridorMap.tsx
- Run unit test suite
<!-- END_AGENT_HANDOFF -->

3. Open Cursor, type "Continue active task", and Cursor resumes the exact next step without missing a beat!

⚡ 97% Context Token Savings via AST Indexing

Instead of dumping full files into the LLM prompt:

💻 Local Development & Web Visualizer

# Clone the repository
git clone https://github.com/shahrukh-hack/vibe-memory.git
cd vibe-memory

# Install dependencies
npm install

# Start local visualizer
npm run dev

🤝 Part of the Vibe Coding Ecosystem

• 🪄 Vibe Superkit: The Anti-AI Slop & High-Taste Design Engine (UI/UX, Spring Physics, Design Tokens).

• 🧠 Vibe Memory: Universal Long-Term Agent Memory & AST Codebase Intelligence.

👤 Author

Created with intention by Yogeshkumar Patel • Adelaide, Australia 🇦🇺

• LinkedIn: https://www.linkedin.com/in/yogeshkumar-ai/

• GitHub: @shahrukh-hack

📄 License

MIT License © 2026 Yogeshkumar Patel