Why AI stories ignore your corrections, and what makes a fix hold

The two-turns-later problem

Anyone who has played a long AI story knows this loop. It is not the model being stubborn, and it is not a bug in the usual sense. It is what happens when a correction is just more text.

Why a fix is outnumbered

In most AI story systems, the story is its own memory. The model reads what came before, sometimes with a summary on top, and works out what is true. Your correction joins that pile as one more line.

The trouble is everything else in the pile. The wrong detail is still written in the chapters before your fix. It may be baked into a summary written before you corrected it. It may sit in notes you gave the model, which compete with everything else for its attention. When the model reconciles all of that, the version that appears most often tends to win.

A correction that has to win an argument every turn will eventually lose one.

This is also why a bigger context window does not fix it. A larger window keeps more of the wrong version in view, too.

What players do today

Players are not passive about it. In a reading of public posts from AI story communities that we ran this summer, about a third of the people who described this problem also described a workaround they had built themselves. The same few come up again and again:

• Regenerate until it passes. Ask for the passage again until the detail comes out right. It fixes one turn, and often costs a turn or a credit.
• Edit the output by hand. Precise, but the model will read its earlier mistake again next turn, wherever else it is written.
• Keep a notes file. Write down what must stay true and paste it back in. It works, and it turns you into the story's archivist.
• Repeat the fact. Mention the green eyes every few turns so the model keeps seeing them. It works, and it bends your writing around the tool.

Each of these is a player doing a job the story should be doing.

What a fix that holds takes

The broad shape is simple to say and hard to build. A fact needs one place where it is true. A correction has to replace the fact in that place, rather than add a line that disagrees with it. The story has to read that place before it writes, so the next chapter starts from the corrected world instead of guessing it again from prose. And making the correction has to be free, or players stop making it.

That is the whole idea behind keeping a record instead of a transcript, which we wrote about in why AI stories forget.

On Nostyss

Here is what is true on Nostyss today, and nothing more. Retry, undo and edit never count against your turns. The story keeps a record of what has happened, and a fact you state in the story is written into that record, not only into the prose. You can see what the story holds in its Memory panel.

Whether a correction then holds for good is exactly what our own tests measure, and it does not hold every time. To measure it on any app, ours included, use the correction test, the third of our five memory tests.

What still breaks

• Deep facts. A detail woven into many scenes is harder to correct than one mentioned once.
• What the story inferred. When the story drew a conclusion from the wrong version, correcting the fact does not always correct the conclusion.
• Your attention. Even when a fix is free, noticing the mistake is still your job. We would like it not to be.

See whether it actually remembers

Ten worlds you can start without an account. Play a few chapters, contradict something you established earlier, and watch what the world does about it. That test is the honest way to judge any of this.

Keep reading