---

⚡ What is DriftBreak?

DriftBreak helps prevent cognitive context-drift and oversized chat histories in long-running local LLM sessions. It extracts and structures established facts, active constraints, and project terminology (State + Glossary) from your conversation log into a lean Recovery Payload.

You load the generated payload into your next session. By reducing context length, the inference engine requires significantly less KV-Cache memory, while model weights remain loaded in GPU VRAM.

Core Architecture

• 100% Offline & Localhost-Only: Operates strictly on loopback (127.0.0.1). Zero telemetry, zero external network calls.
• Lean Recovery Payload: Compresses history into a structured context object (State + Glossary + optional recent turns). Less context length means reduced KV-cache allocation without cold model unloads.
• Chronological Session Vault: Stores raw chat backups, cumulative state histories, and glossaries inside sequential directories (sessions/00001/, sessions/00002/).
• Deterministic Checkpoint Modes:
  • keep 0: Complete reset (empty payload, new session folder; baseline glossary archived to disk).
  • keep 1–15: Active payload injection combining verified State + cumulative Glossary + specified recent turns.
• Atomic Persistence: Safe temp-file writes prevent JSON corruption during execution or user interrupts.

---

🚀 Quick Start (For End-Users)

1. Download DriftBreak.exe from the button above.
2. Place it in any folder and paste your conversation log into session_input.txt.
3. Run DriftBreak.exe and select your local backend node.
4. Use the generated pruned_context_payload.json inside your session directory to resume execution with structured continuity (State + Glossary + recent turns).

---

🛠️ Developer Setup (From Source)

# Clone the repository
git clone https://github.com/Recursive-Logic-Core/DriftBreak.git
cd DriftBreak

# Install requirements
pip install requests

# Run application
python DriftBreak.py