SEP-Memory: Memory-Augmented Explainable Stock Predictions

SEP-Memory is a research and engineering project that integrates the Summarize–Explain–Predict (SEP) forecasting framework with a multi-layer memory system to improve both the accuracy and interpretability of stock predictions.

The system enhances predictive modeling by combining LLM-driven explanations with structured memory layers, reinforcement learning, and efficient retrieval.

---

✨ Features

🧠 Multi-Layer Memory Architecture

• Short-Term Memory: Daily summaries of news, tweets, and stock movements
• Mid-Term Memory: Refined explanations and self-reflections for correction
• Long-Term Memory: Consolidated high-reward patterns and signals
• Reflection Memory: Error cases and failed predictions for targeted retraining

🔍 Memory-Augmented Explainability

• Retrieve relevant past insights with FAISS-powered embeddings
• Inject memory context into prompts for more consistent and reliable explanations
• Enable self-reflective correction loops in explanations

♻️ End-to-End Reinforcement Loop

1. Summarize → Ingest daily market data & generate structured summaries
2. Explain → LLM produces reasoning → reflection step refines explanations
3. Predict → GRPO policy generates trade signals
4. Reinforce → Rewards from real price movements written back into memory layers

⏫ Automated Memory Promotion ("Jump")

• Important knowledge is automatically promoted:
  short → mid → long-term memory
• Low-value or stale knowledge is pruned

---

🛠️ Tech Stack

• LLM Backbone: Transformers + PEFT (LoRA, 4-bit QLoRA)
• Reinforcement Learning: GRPO with reward models
• Memory System: Custom MemoryDB + BrainDB with multi-layer storage
• Retrieval: OpenAI embeddings + FAISS for sub-second lookup
• Training Data: Daily financial news + social media streams

---

📂 Workflow

flowchart TD
    A[Market Data] --> B[Summarize]
    B --> C[Explain v1]
    C --> D[Self-Reflection with Memory]
    D --> E[Explain v2]
    E --> F[Predict with GRPO Agent]
    F -->|Rewards| G[Update Long-Term & Reflection Memory]
    G -->|Promote/Prune| B

Loading