Multi-Turn LLMs: The Costly Memory Illusion

🔥 What happened A researcher lays bare that LLMs have no real memory. Every conversation is rebuilt client-side – and costs explode quadratically with each round. 💡 Why it matters With full replay, input tokens grow quadratically: 10 rounds cost up to 55x more tokens than a single request, not 10x. Caching and truncation mask this, but they shift the failure modes. ⚡ Our take Selling multi-turn as a feature is selling an illusion. The real innovation is in state management – not in the model.