The mature strategy
Real context work looks like a library, not a landfill. You keep an index of what exists, load only what the current task needs, and unload or summarize when the material stops being active. Even if a million tokens fit, you usually should not fill them.
Working set, not corpus
Operating systems do not keep every file in CPU cache. They page in the active working set. AI workflows should do the same: pointer to everything, full text only for what the current decision needs. The pointer is cheap (a path + summary + relevance hint); the full text is expensive in tokens, attention, and noise.
Index lifecycle
The index itself is a stable artifact: list of files, summaries, freshness, when to load. It belongs in the spine or in a project-instruction file the model reads at session start. If the index changes every turn, it is not an index — it is a working set.