Why decay is the wrong model for agent memory facts (and what write-time supersession actually looks like)

.:: Been going back and forth this week on a specific failure mode in agent memory: treating fact staleness like relevance decay. Most memory stacks I've seen (including early versions of ours) embed everything the same way and age it with one decay curve. That works fine for episodic stuff: a Slack message, a one-off event, nobody cares about the exact wording three weeks later. It falls apart for facts and preferences, because those don't fade gradually. A customer's account tier isn't 60% true six months after it changed. It's either still true or it's wrong, full stop. The fix that keeps showing up across different implementations: split memory by type at write time (episodic / semantic / procedural is a reasonable starting split) and run semantic facts through a write-time contradiction check instead of a decay function. Mem0 published a concrete version of this a few months back. Every new fact goes through one of four operations (ADD/UPDATE/DELETE/NOOP) compared against existing similar memories, so contradictions get caught on the way in rather than sitting in the index until retrieval surfaces both versions and the model picks one at random. The part that's easy to underestimate: most memory write-ups focus on retrieval (rerankers, hybrid search, chunking). The write path is where this actually breaks in production. You don't notice the missing supersession check in week one; you notice it three to six months in when an agent confidently states two different things about the same customer in the same conversation. My take: decay and supersession aren't competing strategies, they're for different data. TTL/decay for episodic noise, write-time contradiction checks for facts. Treating one as a substitute for the other is where the "70% confident about a plan tier that changed a week ago" bug comes from. Curious how others are handling this. Are you doing write-time contradiction checks, or relying on recency/decay and hoping it's good enough? And if you've been bitten by the "stacked contradiction" failure, what was the bug report that made you fix it?