AI Contexts That Never Forget - How to give an artificial intelligence a world it keeps

Abstract

A frontier model writes working code and solves competition mathematics. Ask the same model to reconcile three hundred invoices and it loses track of which ones it already processed. A spreadsheet has done that task perfectly since 1985. This article locates the difference, measures it, and names the one operation that closes it. Four reasons this is worth your time. The failure is everywhere and the cause was unnamed. Duplicate actions, stale numbers, a plan built on a fact that changed an hour ago — the industry treats these as separate bugs. They are one mechanism, and this article names it. The field is working on the wrong variable. Almost every 2025–2026 remedy is organised around context length — compress, summarise, evict, fold. This article shows, with the length effect switched off entirely, that length is a bystander. Version multiplicity is the variable. The fix is small, cheap, and measurable. Three operations. The first buys 92% of the available gain and costs less than the baseline it beats. Two standard rebuttals are already dead in the literature. A stronger model leaves the gap open — accuracy falls 92% to 77% on a frontier model, and the gap persists across scale. More memory yields zero recovery: 28% → 28% as memory grows proportionally. The design was specified in 1979 and set aside for tractability. Doyle’s truth maintenance system tracked why each held claim stands, so retracting a premise recomputed what depended on it. Both objections that retired it are engineering, and both are cheaper now. It can be wrong, and it says how. Six tests were committed before any run. Five ran. One refuted the author’s own explanation of the mechanism. The last step is the largest. Repairing the record drives superseded reads to zero and leaves task success flat. Carrying the retirement one step further — into the policy derived from the fact — moves success from 0.782 to 1.000. What this article is. A short, measured account of one mechanism. The experiment is small, the environment is synthetic, and every number is reproducible from the released code. Zero language models were run: that limit is stated here and again at the end. What it uses. The Architecture of Freedom Intelligence — a framework that asks what a system holds as possible before asking what will happen. Section 4 presents it, defends it against its strongest objections, and marks where its own tests located the boundaries.