does anyone have a reliable trick for keeping llms consistent across long workflows?

i keep running into this thing where the model starts clean but after like 8–10 turns it subtly rewrites rules, forgets earlier constraints, or shifts tone even when the prompt is solid. i tried resetting context and even adding tiny checkpoints, but it still drifts once the thread gets long.

i saw a consistency module in god of prompt that separates “stable rules” from “active task logic,” and it kinda helped, but im curious what patterns other people are using. do u rely on memory summaries, isolated rule blocks, or something else to keep behavior steady across longer workflows?