Inspiration
We were inspired by a common problem in AI assistants: they forget users after every session. Users often need to repeat the same information, preferences, and goals again and again. We wanted to build an AI agent that can remember important information, learn from interactions, and provide a more personalized experience over time.
What it does
QMem is a persistent AI memory agent powered by Qwen. It remembers user preferences, goals, and important conversations across multiple sessions. The agent retrieves relevant memories when needed and uses them to generate personalized responses. It also includes a smart forgetting system that removes outdated or less useful memories to keep the memory efficient and relevant.
How we built it
We built the frontend using React.js and Tailwind CSS for a clean and responsive user interface. The backend was developed using Flask and Python. User data and memory metadata are stored in MySQL, while ChromaDB is used for semantic memory retrieval. Sentence Transformers generate embeddings for memory search, and the Qwen API powers the reasoning and response generation. The application is deployed on Alibaba Cloud.
Challenges we ran into
One of the biggest challenges was designing an efficient memory system that could retrieve relevant information without overwhelming the model with unnecessary context. We also faced challenges in ranking memories by importance, implementing memory decay for outdated information, and ensuring fast retrieval while maintaining response quality.
Accomplishments that we're proud of
Built a working persistent memory system. Implemented intelligent memory retrieval using vector search. Created a smart forgetting mechanism for outdated memories. Integrated Qwen for personalized AI responses. Successfully deployed the application on Alibaba Cloud. Developed a user-friendly interface for interacting with the memory agent.
What we learned
Through this project, we learned how memory systems can significantly improve AI interactions. We gained hands-on experience with vector databases, semantic search, embedding models, cloud deployment, and integrating large language models into real-world applications. We also learned the importance of balancing memory retention and forgetting for long-term AI performance.
What's next for Untitled
In the future, we plan to add advanced memory visualization, multi-user support, voice interactions, memory editing controls, and more sophisticated memory ranking algorithms. We also want to improve personalization by allowing the agent to adapt its behavior based on long-term user preferences and goals.