Each time an LLM generates, it can only read a bounded input (tens of thousands to a million-plus tokens). The longer the conversation, the more likely earlier content gets pushed out of the window — which is why a bare chatbot starts forgetting things in long chats.
The fix isn't an infinitely bigger window (cost and speed forbid it), but a persistent storage layer outside the window that retrieves relevant content back in on demand — which is exactly what long-term memory and RAG do.
Try it yourself in Enclave
Open it in your browser — no credit card, no install.