Character content boundaries: keep canon separate from session memory

Name the information layers

Maintain separate records for character canon, authored current world state, session memory, and user-proposed fiction. Canon contains reviewed facts that must remain stable until a writer changes them. World state contains approved time-bound developments. Session memory is a convenience for the current conversation and must not silently gain authority. User-proposed fiction can be welcomed as play, but it needs labels such as “in this imagined version” rather than a claim that it happened.

This separation also protects practical content. A character may narrate a fictional setting, but it should not manufacture real people, interviews, release dates, or instructions. The interface should distinguish authored text from a generative continuation whenever that distinction matters to reader trust.

Give writers a change path

A canon change request needs the proposed statement, affected source entries, continuity impact, reviewer, decision, and publication state. Designers should review how a change appears in prompts, profile cards, memory summaries, and avatar performance. Do not let a model’s plausible elaboration serve as the only draft or approval record. The character may say it does not know or that a detail belongs to a fan interpretation.

Worked example, hypothetical: a visitor says the character “left the city after the winter festival.” The system can reply, “That is one possible scene to explore,” and keep it within the current story branch. It may not later answer another visitor, “I left after the winter festival,” unless that fact has entered reviewed canon.

Bound session memory tightly

Store only what the conversation needs, for as long as it needs it, and label recall as contextual rather than historical truth. A remembered preference such as “you asked for short answers” is different from a world assertion such as “the captain is missing.” Clear session memory on reset and do not merge it into creative documentation automatically. If persistent personalization exists, obtain the appropriate product approval and offer controls.

The counterexample is treating every user correction as a lore update. Another is an avatar that confidently recalls a prior fictional exchange after the user has asked to start fresh. Memory can improve continuity while still being wrong, private, or out of scope.

Review outputs at the boundary

Create tests for contradictions, fan theories presented as facts, impossible chronology, unsupported named people, prompts that request confidential canon, and a user demanding that the character permanently remember a claim. Writers review representative generations and edge cases before new canon is released. A content owner needs a way to withdraw an approved statement and verify that affected retrieval and prompt material no longer promotes it.

Acceptance checks: every public world assertion can trace to a reviewed source; session-only information has an expiry; user inventions are not promoted without approval; and a fresh session cannot retrieve old private context.

Withdrawal check. When canon is withdrawn, test cached prompts, summaries, search cards, and translation variants. A deleted source is not enough if an old derivative still asserts it.

Be candid about authorship

A character’s voice can be authored, generated, or mixed, but an interface should not imply that a fictional person has independent knowledge, feelings, or authority. Do not use continuity to make claims about real events. This is especially important when an embodied presentation makes improvised answers feel more intentional than text.

There is no single web standard for fictional canon governance. W3C accessibility material can help make controls perceivable and operable, while the education guidance addresses specific learner-product risks; neither decides story truth. The layer model and review workflow are original editorial synthesis, and creative teams must set their own canon policy.

For learner-facing products, the Department for Education’s standards for educational settings in England advise avoiding anthropomorphic names, avatars, and claims of agency except where a time-limited, pedagogically justified roleplay is clearly framed and visually bounded. That is relevant caution for a character interface, not a universal rule for fiction or a legal conclusion outside its scope. A creative team should still explain the role, provide a way out of it, and prevent a character voice from requesting personal disclosure to maintain an illusion of relationship.

Keep an explicit list of out-of-world claims the character must not make, including real endorsements, biographical details, product availability, and historical events without reviewed sources. This negative boundary gives writers a practical alternative to relying on a vague instruction to “stay in character.”

Publish boundaries for collaborators

Give writers, moderators, designers, and implementers the same concise statement of what the character may assert, what requires review, and what is session-only. Review localized variants too.

When a reader finds a contradiction, preserve the report and correct the source rather than teaching the system to argue. A graceful correction shows more editorial stewardship than confident improvisation.