The missing primitive in agent memory is not retrieval. It is expiry.

Almost every agent memory system I see is built to remember more and find it faster. Nothing in any of them knows how to stop believing something. A note that says "we tried X, it did not work" is written once and then never expires. Nothing in the system changes on the day it stops being true. It just sits there, gets retrieved, and reads as more authoritative every time it survives another run. A system that only accumulates gets more confident and less correct at the same time, which is a genuinely nasty combination because the confidence is the part people can see. There is a smaller version of this that shows up constantly. When an agent runs out of context mid task, everyone reaches for a handoff note, and handoff notes work strangely well. Worth asking why, because the agent had far more information available than the note contains. They work because they are lossy in the right direction. A note keeps the decision and throws away the reasoning that produced it. The raw context does the opposite: it keeps everything, including the three wrong turns that got corrected later, and restoring it restores the confusion along with the conclusion. So decisions and context are two different objects and most memory work treats them as one. A decision is "we went with X over Z because Y". Context is the forty thousand tokens of finding that out. Only one of those is worth carrying forward, and it is not the big one. What I actually want is a memory entry that carries the conditions under which it was true, so it can be invalidated by the world changing rather than by someone remembering to go and delete it. Nobody is asking a knowledge base to forget. Curious whether anyone here has shipped expiry properly, and what triggered it. Time, contradiction, or someone noticing by hand?