Three signals you've crossed from prompt issues to systemic memory pollution:
1.Behavior drifts without prompt changes (memories are accumulating contradictions).
2.Failures aren't reproducible in a single call (the bad write happened sessions ago).
3.Failures spread across agents (Agent B confidently repeats Agent A's hallucination).
Scale isn't about total size; it's (Write Frequency × Agent Count × Session Overlap). Pollution can start at just a few hundred writes/day.
The leading indicator: Memory Contradiction Rate. If >15% of new writes conflict with existing memories, you're in systemic territory regardless of scale.