TradeMemory

Inspiration

Every AI trading assistant forgets you the moment you close the tab. You explain your risk tolerance, your goals, your history — every single session. And it never learns whether its own advice was right or wrong. I wanted to build a coach that actually holds itself and you accountable over time.

What it does

TradeMemory tracks every market prediction you make, lets you mark it correct or wrong, and scores your accuracy over time — building a real track record. It detects behavioral patterns like panic-selling and FOMO-buying from your conversations, and references your full history in every response, rather than starting from zero each time.

How I built it

• React + Vite frontend with four tabs: Chat, Profile, Predictions, and Memory Log
• FastAPI + SQLite backend that persists conversations, predictions, behavioral signals, and risk profiles
• Qwen Cloud (qwen-plus) as the reasoning engine, with the full memory context injected into every request so responses are always personalized
• A lightweight keyword-based extraction layer that detects asset mentions, prediction language, and emotional/behavioral signals from each message

Challenges I ran into

Balancing how much memory context to inject into each prompt without blowing past token limits was tricky — I ended up weighting memories by importance score so the most relevant behavioral insights are always prioritized.

What I learned

Building the behavioral pattern detection taught me that genuine personalization requires more than remembering facts — it requires building a model of how someone behaves over time, not just what they said once.

What's next

Expanding the behavioral model with more nuanced pattern detection, and adding real-time market data integration so the coach's reasoning is grounded in live prices, not just conversation history.

Built With

• fastapi
• javascript
• python
• qwen-cloud
• react
• sqlite
• vite