The examples below are historical play records from different creator groups and setups. They illustrate observable behavior, not a controlled memory benchmark or the reliability of today's models.
Research checked October 28, 2026, Beijing time. Read the four-source record, exact excerpts, dates and limits. The original play was performed by the source creators, not by this site’s operator.
Did the story reverse a change that had already happened?
In one AI Dungeon detective sequence, the player breaks a character's phone. The narration still has that character answer a call. The player then burns the phone, but another answered call follows. The useful evidence is the sequence: the destructive action appears in the narration before the object is used again. That is a visible continuity conflict. It does not reveal whether the underlying cause was omitted context, a retrieved detail or a generated continuation overriding the scene.
The same historical session supplies a different outcome when the disputed fact is the world's past. A visitor says the nuclear war has not happened, despite the post-apocalyptic premise. Repeated explanations do not settle the disagreement. Sending the visitor outside to examine the damage produces an acknowledgement. Direct correction and a new narrated inspection have different results in this scene. That makes inspection a possible way to continue this fiction, not a universal command for repairing stored memory.
Was the intended action ever established as a fact?
In a separate AI Dungeon peasant scene, the player tries to wrestle a knife from a man's hands. The continuation instead puts the knife in the player character's mouth and then throws it away. The requested disarm is not what the reply narrates. Calling the next turn a forgotten successful disarm would assume an event the story never actually established. First record the difference between what was requested and what the continuation committed to.
An unexpected action can also be welcome. In that session's knight adventure, asking a dragon to talk transforms the hostile encounter into a companion-like exchange and introduces a wizard as another problem. The changed direction is part of the appeal there. The knife scene and dragon scene therefore support no simple rule that deviation is always useful or always broken. The player's intended outcome and the accepted continuation both matter.
Are you attempting an action or directly changing the story?
In a historical NovelAI sewer encounter, repeated attempts to remove an antagonist do not settle the exchange. The players discuss switching from an attempted action to a direct statement that the character dies and does not return. The next continuation leaves the protagonist alone. That is a change of intervention: asking the story to resolve an attempt and editing what is already true are not interchangeable instructions.
NovelAI's Text Adventure documentation distinguishes action and dialogue input from direct story editing. Those terms help explain the difference; they do not prove that a direct edit updates every other memory representation. In the same historical record, written endings can still be followed by renewed narration. One locally useful intervention is not a general guarantee that a scene stays closed.
Does a repeated action mean the AI forgot it worked?
A human-run D&D session with AI player characters provides a useful boundary. During a golem fight, an AI character lays a slowing slime trail. Later, it proposes another trail around the same enemy. The human game master points out that this has already been done, while continuing to describe the enemy as constrained. The repeated proposal and the retained consequence coexist. Repetition alone does not establish that the earlier effect vanished from the game state.
Another moment in that encounter separates proposed success from adjudicated success. An AI character stuck in a well describes assembling an improvised elevator and scrambling out. The human game master rejects the successful exit and keeps the character at the bottom. A later flight-based attempt is accepted. Here the human controls which proposed action becomes a result. It is not evidence of an autonomous AI narrator reliably maintaining that boundary, but it shows why an action declaration cannot simply be counted as a completed state change.
What do these examples actually let you conclude?
They let you distinguish a contradiction from an unfulfilled request, a change of editing mode and a repeated proposal. They do not expose the full context supplied to the models, the contents of hidden summaries or a common memory limit. The sessions use different scenes and arrangements; the human-run encounter is especially different from an AI-generated narrator. No shared failure rate or product ranking follows from them.
For a case where the full source text reached the prompt but the reply still changed facts, read our Story Card migration test: exact transfer versus factual consistency.
| Observed problem | Compare these two things | Limit of the evidence |
|---|---|---|
| A destroyed phone answers again | Narrated destruction and the later use | Shows conflicting continuity, not its internal cause |
| A disarm becomes a different action | Requested command and actual continuation | Does not establish a successful disarm was later forgotten |
| A character remains in an unwanted scene | Attempted action and direct story intervention | A local edit does not certify all future memory |
| A tactic is proposed again | The repeated proposal and the effect still in force | Repetition is not proof the earlier effect disappeared |
| A proposed escape is treated as completed | The character's declaration and the adjudicated result | Human adjudication is not an autonomous-memory test |
Take the first disputed reply and write down the earlier fact it contradicts; if you cannot name that fact, start with the command's interpretation rather than a memory diagnosis.
