One of the central ideas in First Person Viewpoint is that the world should remember what happened, not merely repeat the last few messages back to the model.
A long context window can help an AI narrator see more text, but text history and world state are not the same thing.
Story text versus state
Imagine a scene where you find a rusty key, give it to Mara, and later use the unlocked bridge to leave the valley.
The prose describing those events is useful history. But the application also needs facts it can rely on later:
- you no longer have the key;
- Mara received it;
- Mara's trust changed;
- the bridge became available;
- your location may now be different.
Those facts should not depend on the narrator rediscovering them inside thousands of words of old prose.
Persistent consequences
FPV treats important changes as structured parts of the world. That allows later scenes to be generated from both the narrative and the current state.
Still experimental
There is no perfect memory model for open-ended fiction. Some details should be permanent, some should fade, and some should matter only to a specific character or location.
That is one of the areas I am continuing to experiment with as the project evolves.


