Thanks! That is the challenge.

Each memory type handles chaos differently: working memory overwrites (no pileup, latest state wins), semantic memory deduplicates automatically via cos similarity so dup knowledge merges instead of accumulating, and episodic memory is append-only but scoped by agent + session + time range so you can always filter to exactly what you need. The mess stays queryable. Want to try it?