Introduction
Most automated narrative systems focus on the logical "plumbing" of a story—ensuring that Event A leads to Event B and that characters act according to their goals. While this produces coherent plots, it often ignores the most critical component of the narrative experience: the reader’s mind. Humans do not just track logic; we build dynamic "situation models" where some details remain vivid while others fade into the background.
If an AI storyteller cannot predict what a reader is likely to remember at any given moment, it cannot effectively create suspense, plant "Chekhov’s Guns," or manage complex plot twists. This paper addresses this gap by introducing INDEXTER, a computational model that bridges automated story generation and cognitive psychology to provide a formal representation of reader memory.
Theoretical Framework
INDEXTER builds upon Partial-Order Causal Link (POCL) planning, a standard AI method for generating sequences of actions. The authors extend this framework by integrating the Event-Indexing Model (EIM), a theory from cognitive psychology which posits that readers track five specific dimensions for every narrative event:
- Time: When the event occurred.
- Space: Where the event took place.
- Causality: How the event relates to prior occurrences.
- Intentionality: How the event relates to character goals.
- Entities: Which characters or objects were involved.
By mapping these dimensions directly onto plan steps, INDEXTER allows an AI planner to reason about memory using the same formal logic it uses to reason about physical actions.
Quantifying Recallability
The core contribution of the paper is the formalization of a "Recallability Formula." This quantitative measure treats memory as a relative, dimensionless quantity rather than a binary state of "remembered" or "forgotten."
The model defines recallability as the ratio of the overlap between a "memory cue" (the current event) and a specific past event, compared to the total overlap across all events in the story’s chronology:
$$recallability(e^*_i, q, C) = \frac{overlap(e^*_i, q)}{\sum_{e_i \in C} overlap(e_i, q)}$$
Essentially, if a new event shares multiple indices—such as the same character, location, and goal—with a previous event, that older event becomes significantly more "recallable" in the reader's mind. This allows the system to rank every past event by its likely cognitive availability at any point $q$ in the narrative.
Narrative as Memory Management
The central insight of this work is that storytelling is an act of managing a reader's cognitive load. Authors do not simply provide information; they manipulate the availability of past events for cognitive processing. By using INDEXTER, a generative system can move beyond simple plot consistency to intentional memory management. For example, a system could ensure a reader remembers a specific clue right before a major reveal or intentionally suppress a detail to heighten the impact of a surprise.
Limitations and Future Work
As a foundational architectural paper, INDEXTER currently faces several challenges:
- Empirical Validation: The model is currently a mathematical framework and has not yet been tested against human reading data to verify if the recallability scores align with actual human memory performance.
- The Consistency Assumption: The current model assumes a logically consistent chronology. However, human readers often hold inconsistent or "broken" models of a story world. Future iterations will need to handle these logical contradictions.
- Generative Application: The ultimate goal is to transition from analyzing memory to generating stories designed to achieve specific "mental configurations" in the reader.
Summary updated