Agents Never Remember What Isn't There
Most agent memory systems have a small operation table. Write a note. Read it back. Supersede it with a newer value. Delete it. The store is a record of what an agent found: a definition, a call site, a config value, a decision, a gotcha. Every design decision in the system assumes the note names something that exists.
Almost none of these systems can store the other kind of finding: that something does not exist.
There is no retry configuration in this service. Nothing calls this function. This repository has no integration test layer. That flag was never wired up. Each of those is a real finding, each is expensive to establish, and each is thrown away at the end of the session that established it. The next session pays for it again from zero.
The argument of this post is that the gap is structural. An absence cannot be stored the way a presence is stored, because its truth conditions are inverted in a way that breaks every mechanism a memory system uses to keep notes honest. Closing the gap requires a primitive nobody has built, and the rest of the post is about what that primitive would have to look like, why three reasonable alternatives fail, and what has to be true of the layer below the memory system for the primitive to work at all.
