Tiered Memory and Just in Time Loading

Suppose Desk has 30,000 tokens of stored knowledge about how the store's cases tend to go. A 10-step run with no memory sends 141,000 tokens. If your code places all 30,000 tokens in every call, the same run sends 441,000.

Now suppose the run needs only 2 of the 40 topics in that store, and your code keeps a 1,200-token list of topics in the window instead. The model asks for the 2 topics it needs, which takes 2 more steps. That run sends 189,450 tokens, about 43 percent of the larger figure