Context churn: why Grok bot usage tanks once context passes 25-30% of the window

alexcovo_eth · x · 2026-10-22

Replying to a usage-testing thread, alexcovoeth explains why users complain that "it was so good until it wasn't": context churn. The bot reads its full context every turn—great for continuity, terrible for usage—and once context hits 25-30% of the window, consumption drops off a cliff.

His fix: retire bots regularly and manage context bloat proactively. He also criticizes Grok bots for showing neither usage nor context window size nor offering any compacting/pruning controls—a real gap in an otherwise excellent product. In his view usage hasn't actually gotten worse; it's a skill issue.

Original post →

Context churn: why Grok bot usage tanks once context passes 25-30% of the window

More from Models