I'm assuming that they have implemented it via a MemGPT-like approach, which doesn't clog the context window. The main pre-requisite for doing that is having good function calling, where OpenAI currently is significantly in the lead.

I've been finding with these large context windows that context window length is no longer the bottleneck for me — the LLM will start to hallucinate / fail to find the stuff I want from the text long before I hit the context window limit.