Should a Fictional AI Character Remember Facts or Feelings?
For a fictional AI character to stay consistent, its memory should preserve what happened and keep the character’s reactions alongside it as a separate record. Treat confirmed story facts as anchors; label feelings as the character’s own stated experience, and leave interpretations uncertain unless the story confirms them. This lets the character respond with continuity without turning a guess about another person into canon.
Why facts and feelings need different memory fields
A fact describes an event or state in the story: Mara left the brass key on the windowsill before the storm. A character reaction describes that character’s experience: Mara later said she felt relieved when the key was found. An interpretation goes further: perhaps she feared that someone would discover what the key opened. The first two may be directly established in a scene; the third is a hypothesis unless a later scene confirms it.
These distinctions matter because a model can turn a plausible explanation into a remembered event when it summarizes a long conversation. Research on generative agents describes a memory stream from which relevant records are retrieved, alongside a process that synthesizes higher-level reflections. The authors report that observation, planning, and reflection each contributed to believability in their experimental setting. For character writing, that architecture is a useful prompt to retain both scenes and summaries, while marking summaries as derived interpretations rather than original events. Generative Agents: Interactive Simulacra of Human Behavior
What should count as a confirmed story fact?
Store details that the narrative explicitly establishes or that the user confirms as canon: who was present, what was said or done, where an object was left, and which events followed others. Keep the source nearby. A short record might say: “During the station scene, Ivo handed Mara the brass key. Source: user-confirmed scene summary.” If a character says something about their own reaction, record the attribution too: “Mara said she felt relieved when Ivo returned. Source: Mara’s dialogue in the station scene.”
This is a practical application of provenance: W3C’s PROV model describes information about entities, activities, and agents involved in producing something, so its history can help readers assess reliability. It is not a character-writing rule, but it suggests a useful design choice: preserve where a memory came from and how it was created. W3C PROV-O: The PROV Ontology
How should the character remember its own feelings?
Keep reactions in the character’s voice and attach them to the scene or statement that supports them. “I was embarrassed when I dropped the map” is stronger evidence of the character’s stated feeling than “The character looked away,” which is an observable action that admits multiple readings. A scene can hold both: “Mara looked away after the map fell. She did not explain why.” If Mara later says she was embarrassed, add that as a later statement; do not silently rewrite the earlier observation as proof.
Feelings can change. A useful record can preserve the time and context of each reaction instead of flattening them into a permanent trait. “Mara said she was relieved when the key was found” and “Mara later said she was uneasy about who had kept it” can both remain true. A 2025 study of episodic memory for dialogue separates memory fragments into factual, experiential, and subjective information, and represents relations among items across sessions. Its results concern conversational AI, not fictional-character quality; still, the categories offer a grounded starting point for keeping a scene, an experience, and an interpretation distinct. Pre-Storage Reasoning for Episodic Memory
How can uncertain interpretations stay useful without becoming canon?
Store an interpretation as a possibility with its evidence, owner, and status. For example: “Possible reading: Mara may have been worried about the key. Basis: she looked away and changed the subject. Status: unconfirmed.” This gives future writing something to revisit, while making clear that no scene has established her motive. If the story later confirms or rejects the reading, update its status and retain the earlier basis where useful.
Do not infer a real user’s emotional state from fictional dialogue, a role-play choice, or an ambiguous action. Keep statements about the real participant separate from the character’s story; only carry over information about the participant if they state it plainly and confirm it is relevant. This boundary is a design recommendation for the task, not a claim that an emotion can be reliably detected from a particular phrase or gesture. In dialogue research, uncertainty can be represented as a system outcome rather than forcing a best guess: one ACL paper studies uncertainty classification and a “none of the above” option in response retrieval. That supports the narrower design lesson that a memory or response system can preserve uncertainty instead of pretending every prompt has a settled answer. “None of the Above”: Measure Uncertainty in Dialog Response Retrieval
A practical memory record for scenes and reactions
A compact record can use separate fields for the event, the source, the character’s own stated reaction, and any unsettled interpretation. This schema is an editorial tool, not a standard required by the cited research:
Field: Event; What to record: Confirmed action or state; Example: Ivo handed Mara the brass key at the station.
Field: Source; What to record: Where the information came from; Example: User-confirmed scene summary.
Field: Character statement; What to record: The character’s own reported reaction; Example: Mara said she felt relieved when it was returned.
Field: Observation; What to record: Visible or audible detail without assumed motive; Example: Mara looked away before answering.
Field: Interpretation; What to record: A possible reading, marked unconfirmed; Example: She may have been uneasy; the scene does not establish why.
Field: Status; What to record: Canon, stated, inferred, or unresolved; Example: Event: canon; reaction: stated; motive: unresolved.
That separation resembles schema-based dialogue state tracking, where a system fills values into predefined slots rather than treating all conversation text as one undifferentiated record. The cited study concerns task-oriented dialogue, so applying the idea to story memory is an analogy; the practical benefit here is simply that each kind of information has a designated place. Dialogue State Tracking with a Language Model using Schema-Driven Prompting
How to review and update a character’s memory
After a scene, first extract only events the scene or user clearly confirms. Then add any reactions the character explicitly states, preserving the scene and speaker. Finally, record interpretations separately with a brief reason and an unresolved label. During retrieval, prefer the original scene or statement when it is available; use a summary as a pointer, not as stronger evidence than the source it summarizes.
When new writing changes the story, update the relevant fact deliberately. If two accounts conflict, keep the conflict visible until the user resolves it or the narrative establishes what happened. Do not settle the conflict by choosing whichever account sounds most coherent. W3C’s provenance materials describe recording entities, activities, attribution, and derivation; applying that idea here helps a writer trace a summary back to a scene and distinguish original material from a later interpretation. PROV Overview
A simple decision rule
When deciding what a character should remember, ask: Was this directly established, and who said or showed it? If the answer is clear, preserve the story fact with its source. If the character stated a feeling, preserve it as that character’s reaction in that moment. If the record describes a motive, emotion, or meaning that no one confirmed, label it as an interpretation or leave it open. The result is memory that supports continuity and character voice while allowing the story’s unanswered questions to remain unanswered.