Inspiration
AI assistants often require users to repeat their project details, preferences, and experience level in every conversation. We built TRACE AI to explore how an assistant can retain useful context while letting users decide what it remembers.
What it does
TRACE AI helps users answer questions, plan projects, and solve problems. Users can extract memory suggestions from their messages, review them, and approve the facts they want the assistant to use.
Approved memories remain available when starting a new conversation within the same browser session. Users can remove individual memories or forget them all.
How we built it
We built the application in Python with Streamlit for the interface. Nebius Token Factory generates chat responses and extracts memory
Streamlit session state keeps each visitor’s conversations and memories separate. The application is hosted on Streamlit Community Cloud, with source code managed on GitHub.
Challenges
The main challenges were separating visitor data, distinguishing conversation history from approved memories, and ensuring pending suggestions never enter the assistant’s memory context. We also added request limits, bounded conversation context, and error handling to make the public demo more reliable.
Accomplishments
We deployed a working prototype with chat, memory extraction, approval, recall, and deletion. Ten automated regression tests cover key behaviors, including visitor isolation and memory controls.
What we learned
Useful AI memory needs more than storage. Users need clear controls over what is remembered, when it is used, and how it can be removed.
What’s next
The next step is authenticated accounts and persistent storage so approved memories can safely remain available across browser sessions. The current prototype stores memory within the active browser session.
Built With
- factory
- nebius
- python
- python-dotenv
- streamlit
- token
