The write and forget sections are the two that decide whether a memory system helps or slowly poisons itself. Retrieval gets all the attention, but the failure I keep seeing is a memory written from a single throwaway sentence - a user says they prefer dark mode while testing something - and that fact then outranks their actual current preference for months, because nothing ever expires it and every retrieval faithfully surfaces it. Distinguishing episodic from semantic helps here in a very practical way: an episode is true about a moment and should keep its timestamp, while a semantic fact claims to be true now and needs a supersede rule. The other thing worth stating explicitly next to the Mem0 section is that a memory layer changes your debugging story - once answers depend on state accumulated over months, reproducing a bad response means reproducing the memory, not just the prompt.