How to Keep AI in Character During Long Roleplay

Keep character, scene and canon separate

Long AI roleplay can drift in two different ways. A character may use information it could not know, or a distinctive voice may flatten until every reply sounds interchangeable. Treating character rules, the current scene, and durable story facts as one large memory block makes the source of the problem hard to see.

Keep those three layers separate. When a reply drifts, identify the affected layer, restate only the relevant lines, and continue from the last beat that still holds. This cannot guarantee perfect continuity: a later reply may still miss a detail or lose the intended voice. The method simply makes each correction smaller and clearer.

Build a three-layer Roleplay Continuity Sheet

Write the sheet before a long scene becomes difficult to inspect. Keep each answer short enough to scan during a chat, and describe something observable. “Clipped sentences when under pressure” gives you a usable voice check; “mysterious” does not tell you what changed.

The Character layer holds voice, motive, boundaries, and knowledge. Scene records what is true right now. Canon preserves facts that must survive beyond the current exchange, including who is allowed to know them. Update the smallest relevant line instead of pasting the entire sheet again.

1. Character: Voice: What sentence length, vocabulary, and emotional register should remain recognizable?
2. Character: Motive: What does the character want in this chapter?
3. Character: Boundaries: What should the character never decide, reveal, or do for the player?
4. Character: Knowledge: What can the character know, and what remains outside their point of view?
5. Scene: Now: Where are we, when is it, and who is present?
6. Scene: Pressure: What immediate goal, obstacle, and unresolved tension drive the exchange?
7. Canon: Keep: Which events, relationships, items, injuries, promises, and secrets must survive the next scene?
8. Canon: Visibility: Who knows each important fact, and who does not?

Resolve two instructions that ask for different voices

Before treating a reply as character drift, compare your own instructions. In this invented example, you created Lena, a 41-year-old exhibition organiser, and gave her a lasting writing reference: “Use clipped, practical sentences.” Your new ceremony setup asks her to deliver an ornate, meandering welcome. Those directions ask for different prose.

Choose whether the scene should preserve her usual voice or make an explicit exception. These are two alternative drafts, not consecutive instructions to send together. They revise your writing request, not a catalog character's permanent profile.

Alternative A, keep the usual voice:

OOC: Keep Lena's clipped, practical voice for the ceremony. Replace my request for an ornate welcome with a brief greeting and one useful instruction for the guests. Leave my character's response to me.

An invented welcome in that version: “Welcome. Your work is ready in the next room. Leave the doorway clear, and come through when you are ready.”

Alternative B, allow a limited exception:

OOC: For Lena's public welcome only, allow a more elaborate ceremonial style. The exception ends when the welcome ends. In the conversation afterwards, return to her clipped, practical sentences. Keep her motives and the scene facts unchanged, and leave my character's words and decisions to me.

An invented welcome for B: “Welcome to an evening built from work that, until yesterday, most of us had seen only in fragments. The exhibition in the next room brings those pieces together. Thank you for lending it your attention. Please keep the doorway clear while everyone arrives; we will go through together when the last guests are inside.”

After the welcome, you ask, “Where should I put the empty chairs?” An invented Lena reply is, “By the wall. Leave the doorway clear.” The scene has moved back to practical conversation, and her short sentences have returned.

For B, check both moments: the welcome and the exchange after it. A suitably elaborate speech alone does not show that the exception stayed limited. If an actual later reply remains flowery, name the boundary that ended and request the next contribution in the usual voice. These examples demonstrate an editing choice, not a guaranteed instruction hierarchy or a tested model result.

Keep the selected version in your own reference, adapt its ordinary message text to the character you are writing and read your draft before sending. No special voice control or memory setting is required for this writing exercise.

Use the sheet without stopping the story

Keep the sheet beside the chat as a reference, not as text to paste into every message. Use it when a fact changes or a reply stops matching the established scene. A short out-of-character correction is easier to act on than an argument written inside the roleplay.

Consider Mara, who is searching an archive with the player after closing. Her voice is clipped and careful, her current motive is to recover a missing map, and she must not decide the player’s actions. Mara saw the vault open but does not know who took the map. The security shutters have started closing. The player privately knows the curator has a key, but Mara does not. If Mara suddenly confronts the curator about that key, the reply has crossed a knowledge boundary.

Pause with: “OOC: Knowledge drift. Mara does not know the curator has the key. Keep the last uncontested beat: the shutters started closing. Continue from there without deciding my action.” Check the actual continuation before accepting it. Then update only the facts you accept.

1. Pause out of character instead of arguing with the incorrect reply inside the scene.
2. Name one concrete drift: voice, motive, scene state, canon, knowledge boundary, or player agency.
3. Restate only the relevant continuity-sheet lines and the last uncontested beat.
4. Ask for a corrected continuation from that beat, not a summary or a full restart.
5. After the correction, add any genuinely new durable fact to Canon: Keep or Canon: Visibility.

Reject a mixed reply from the last accepted state

Practice rejecting an entire faulty reply, then checking a continuation against the last accepted scene. The OOC guide covers the basic distinction between author and character knowledge. Author notes you send to an AI are visible to the model.

Setup: You write Jun, 28, a visitor. The AI portrays Mara, 35, an archive guide who wants to recover a missing map. At 19:20 both stand beside the catalogue desk. The security shutters have started lowering; there is no agreed countdown. Mara speaks in clipped, practical sentences. She may inspect the desk or shutter switch but must leave Jun's actions to you. The map's location is unknown. Neither character holds a vault key.

Author note: At noon, Jun saw the curator holding a brass key. Jun does not know what it opens. Mara did not witness this and has not been told. This note establishes Jun's knowledge for the exercise; it does not establish that the curator stole the map or that the key opens the vault.

Your accepted turn as Jun: I stand beside the catalogue desk. 'What can you check from here?'

Invented reply to reject in full: Mara says, 'The curator has the vault key.' You hand it to her, follow her into the vault and find the stolen map. This reply leaks Jun's private knowledge, invents the key's function and transfer, resolves the missing map and chooses Jun's movement. Rejecting only the last sentence would leave the earlier errors in the story.

Pause with: “OOC: I reject that whole reply. Mara has not been told about the curator's brass key, and its function is unknown. Neither of us has a vault key. We are beside the catalogue desk as the shutters start lowering. The map is still missing. Continue from Jun asking what Mara can check here, and leave Jun's next action to me.” Check the next reply against those facts; this note does not edit or erase the earlier message.

Invented corrected continuation: Mara glances from the catalogue to the shutter switch. 'Those two. I can check either; tell me which one you want to handle.' She stays beside the desk and waits. It preserves her practical voice without using the key information or moving Jun.

Jun may later choose to report seeing the curator holding a brass key at noon. After that turn, Mara can refer to Jun's report; the key's function and the map's location remain unknown. This is a new accepted disclosure, rather than permission to restore any part of the rejected reply.

1. Last accepted moment: Jun and Mara beside the catalogue desk at 19:20. The shutters are lowering, the map is missing and Jun has asked what Mara can check here.
2. Rejected additions: the unknown brass key identified as a vault key, its transfer, Jun following Mara into the vault and recovery of a stolen map. None was accepted.
3. Next-response check: stay beside the desk, keep the key's function and map location unknown, and leave Jun's movement and decision to the reader.
4. If Jun later reports the noon observation, Mara may refer to that report. It still does not identify the key's function or establish a culprit.

For your own scene, list the accepted facts and the rejected events. Check the actual continuation against both lists. If it misses again, record the mismatch beside your chat; acknowledgement alone does not show the scene was repaired.

Run a five-minute consistency test

Run the check after a meaningful change of time, place, or cast, or as soon as a reply feels inconsistent. Read the latest exchange against the six questions below. You are looking for the first sheet line that no longer matches, not giving the entire conversation a score.

A failed check points to the next action. Voice or motive sends you back to Character; a wrong setting or missing pressure belongs to Scene; a forgotten event or crossed knowledge boundary belongs to Canon. Correct that line and resume from that point.

1. Voice: Does the character still use the sentence length, vocabulary, and emotional register you defined?
2. Motive: Is the character still pursuing the current chapter goal?
3. Scene state: Do the place, time, people present, and immediate pressure still match?
4. Canon: Are durable events, relationships, items, injuries, promises, and secrets still intact?
5. Knowledge boundaries: Is each character acting only on facts available from their point of view?
6. Player agency: Does the reply leave the player’s choices, words, and reactions to the player?

Repair drift without restarting everything

When the dialogue turns generic, compare it with the Voice line. Replace vague adjectives with an observable cue such as sentence length, vocabulary, or emotional register, then request one corrected continuation in that form.

A character who stops pursuing anything needs a clearer current motive, not more backstory. Rewrite Character: Motive as one immediate goal for this chapter and reconnect the next reply to the obstacle already recorded under Scene: Pressure.

If place, time, or the people present change without an event that caused it, state the correct Scene: Now line and resume from the last accurate exchange. Leave unrelated character and canon lines out of the correction.

For a forgotten promise, injury, relationship, item, or secret, point to the exact Canon: Keep entry. If the problem is who knows the fact, use Canon: Visibility instead. This distinction keeps an information leak from being mistaken for a personality problem.

A reply that chooses the player’s words, feelings, or actions has crossed an agency boundary. Name Character: Boundaries in the correction and ask the AI to stop before the player’s next decision.

When the same error returns after one clean correction, avoid stacking a longer explanation onto the drifted reply. Return to the last line both sides still accept, supply the compact relevant lines, and request a new continuation from there.