A default tracking setting in Claude Code can slow down local model performance by ninety percent.

When Claude Code talks to a model, it sends a hidden tracking tag with every message. For local models, this constant tagging ruins their short-term memory, which is called the KV cache. It forces the local AI to re-read your whole project from scratch every single turn. Turning off this tag by setting a specific option to zero makes your local AI run up to ten times faster.