FINMEM Streamlit — Clean Refactor (Daily Trading Agent)

A production-ready, simplified refactor of a FinMem-style daily trading agent with:

• Personas: Secure / Balanced / Risk (dropdown)
• Train/Test modes with full evaluation loop
• Metrics: cumulative return, Sharpe, max drawdown, #trades
• Equity vs. Price plots + actions table
• News sources: Auto / NewsAPI / Alpaca (fallback to none)
• Hierarchical memory scaffolding (short/mid/long + reflections)
• OpenAI LLM wrapper with robust fallback (Responses → Chat Completions)
• Windows-friendly dependencies (no hard FAISS requirement)
• Clear logs & checkpoints under data/runs/<timestamp>/

No secrets are committed. Provide your keys via .env (see .env.template).

---

Quickstart

1. Install (Python 3.10+ recommended):

pip install -r requirements.txt

2. Set environment (create .env from template):

cp .env.template .env
# edit .env to add your keys (OPENAI_API_KEY, NEWSAPI_KEY, ALPACA_API_KEY/SECRET)

3. Run the app:

streamlit run app.py

---

Notes

• Data source: Default price data via yfinance (no API keys). If you already have pickled env data, you can load it from the UI.
• Personas: Map to different risk appetites and prompt styles. You can still edit the system prompt if you wish.
• News: If you select Auto, we use NewsAPI when NEWSAPI_KEY is present; otherwise, try Alpaca if keys exist; else skip news.
• Memory: Uses OpenAI embeddings when available. If not, a deterministic hash-based local embedding kicks in (lower quality but functional).
• LLM: You can type any OpenAI model name (e.g., gpt-4.1, o3, gpt-4o-mini). Wrapper attempts Responses API and falls back to Chat Completions automatically.

---

Project Structure

finmem_streamlit_clean/
├─ app.py
├─ requirements.txt
├─ .env.template
├─ README.md
├─ finmem_core/
│  ├─ __init__.py
│  ├─ agent.py
│  ├─ chat.py
│  ├─ config.py
│  ├─ environment.py
│  ├─ memorydb.py
│  ├─ metrics.py
│  ├─ news.py
│  ├─ plot.py
│  ├─ prompts.py
│  └─ utils.py
└─ data/
   ├─ runs/        # created automatically
   └─ cache/       # optional caches

---

What’s Improved vs. Your Previous Iteration

• Proper Test loop (iterate until done; compute metrics, plot, and table).
• Persona selector with three presets and consistent mapping.
• Optional news source selector (Auto/NewsAPI/Alpaca/None) + article list.
• Config warnings surfaced in UI when present.
• Duplicate utilities removed; a single utils.py provides ensure_dir/stamp_dir.
• OS-friendly requirements (no forced faiss-cpu).
• Logging and checkpoints are saved and linked from the UI.

---

Disclaimers

• This repo is designed for clarity and stability. It is not a perfect reproduction of every research nuance of the original FinMem paper; rather, it keeps the spirit (memory hierarchy + LLM decisioning + personas) while being practical.
• For real trading, add execution logic, slippage/fees, risk controls, and compliance checks.

Enjoy!