Core idea

The key idea is not to create a model with a native billion-token attention window.

Instead:

$$\text{very long cumulative session} \rightarrow \text{context management} \rightarrow \text{limited active context} \rightarrow \text{LLM}$$

So a session may process billions of cumulative tokens over time while only a much smaller amount of information remains resident in the model's active context at any one moment.

In short:

Keep recent information verbatim, fold older information into recoverable summaries, and retrieve historical details only when they become relevant again.

---