Build an Agent with Long-Term, Personalized Memory

---

---

AI Summary

Summary: Demo of a Cooking Application with Memory Feature

Introduction

• Demonstrated a cooking application that mimics chatbot long-term memory features.
• The application learns about the user’s cooking habits, allergies, and preferences.

Application Features

• Extracts information from user conversations about allergies, food preferences, and family attributes.
• Uses this information for meal planning and personalization.

Technical Approach

• The application uses two agents: a conversational agent and a memory agent.
• The memory agent listens to messages and stores relevant information.
• The memory feature is essential for the app to remember user-specific details like dietary restrictions.

Implementation Details

• The memory agent is split into two steps for efficiency:
  1. A low-cost model (memory sentinel) checks if a message contains new information.
  2. A more powerful model (memory manager) processes new information using GPT-4.
• The memory manager uses a custom tool to create, update, or delete memories in a database.

UI Design

• Adopted a top-down message display inspired by Gemini, as opposed to the traditional bottom-up chat interface.

Research and Inspiration

• Inspired by the MGPT project and research paper.
• MGPT combines long-term and short-term memory in chatbots, condensing information into a small context window.

Code and GitHub

• Shared code on GitHub for others to experiment with.
• The code includes the memory sentinel and memory manager setup using Lang Chain and Lang Graph.

Optimization and Future Considerations

• Discussed potential optimizations for cost and efficiency.
• Suggested improvements like vector storage for memories and session-based memory updates.

Conclusion

• The demo showcased a simple yet effective implementation of long-term memory in a cooking application.
• Encouraged viewers to access the code on GitHub and experiment further.

GitHub Repository Access - Link provided for viewers to access the demo code.