https://www.canva.com/design/DAG2KoK203w/duvoIHpRy7zUYRXafYvZlQ/edit?utm_content=DAG2KoK203w&utm_campaign=designshare&utm_medium=link2&utm_source=sharebutton

Mcpmem

Memory with context, time, and portability—built for LLMs that know you.

---

The Problem We're Solving

Context engineering is the biggest bottleneck for LLM adoption. ChatGPT's Memory feature exists, but it's fundamentally broken:

• No control: You mention something once, it haunts you forever—no editing, only deletion
• No structure: Flat memory with no relationships, no time dimension, no organization
• No evolution: Information piles up with equal weight. Facts change, preferences shift, but the memory stays static
• No portability: Locked into one provider's ecosystem

The real issue? Most people don't use complex prompting techniques. They ask simple questions like "reply to this email" or "write this document" and get generic responses that don't fit. Then they spend forever going back and forth trying to fix it—starting from scratch every time.

What's missing? A memory system that understands relationships and time. When you tell your AI "I now prefer Nike over Adidas," it shouldn't just add a new fact—it should update the knowledge graph, deprecate old preferences, and maintain the connection between related information. That's how real memory works.

People need their AI to actually understand them and evolve with them, but there's no good way to build and maintain that ultra-personalized, living context.

---

Our Approach

Portable, visual, temporal memory—no maintenance burden.

1. Temporal Knowledge Graphs (Graphiti + Neo4j): The game-changer. Memories aren't just stored—they form relationships over time.

   • "Claire liked Adidas" → "Claire sold Adidas stock" → "Claire now prefers Nike"
   • Entities connect, preferences evolve, outdated facts expire naturally
   • Ultra-personalization that adapts to you as your life changes

2. MCP (Model Context Protocol): Plug memory into any modern LLM app. Zero lock-in, maximum portability. Your memory graph travels with you.

3. Obsidian-style visualization: Explore your memory as an interactive graph. Search episodes, trace connections, see why the AI knows what it knows.

What you get:

• Automatic memory creation from LLM interactions (chat histories, documents, notes—everything can be stored)
• A living knowledge graph that updates relationships constantly
• Search, edit, delete—full user control
• Transport your memory graph anywhere via MCP

Think "Google Drive for LLM Context"—one intelligent memory system that works with any LLM you prefer, learns from every interaction, and evolves with you.

(Note: Add/delete tools are WIP due to hackathon time constraints, but the core search and temporal graph engine are live.)

---

How to Launch

1. Environment setup

cp .env.example .env
# Fill in required keys (Neo4j, LLM API, etc.)

2. Start services

docker compose build
docker compose up

3. Seed the database

curl -X POST http://localhost:8000/neo4j/seed

4. Explore your memory

• Web UI: http://localhost:5173 – visualize and interact with the graph.
• Test MCP:

  npx @modelcontextprotocol/inspector http://localhost:8000

  • Auth method: SSE
  • URL: http://localhost:8000/mcp/sse
  • Try the search tool!

---

Acknowledgments

Built with 🫩 for Cursor Hackathon Singapore.
Thank you to the organizers for the energy, the community, and the tight deadlines that make magic happen.

—The Mcpmem Team