Long-context retrieval

Every token admitted to a full context window evicts one already there

A field-notes post on production LLM agents argues the context window behaves like a cache under eviction pressure, and that the real work is deciding every turn what earns space and which failed turns to scrub.