I’m looking into persona drift in long-running LLM systems.

Kotodama OS is an attempt to treat this as an external behavior-layer problem, rather than solving it with more prompting or fine-tuning.

I opened a GitHub issue to collect real-world cases where persona drift actually caused problems, and to discuss possible integration patterns.