
Context Decay: Why LLMs Forget Your Instructions in Long Conversations
You give a large language model a clear instruction at the start of a long document — "Respond only in JSON" or "Never reveal the customer's account number" — and 20,000 tokens later, it quietly ignores you. The rules you…

















