> For the complete documentation index, see [llms.txt](https://docs.seaart.ai/guide-1/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.seaart.ai/guide-1/seaart-ai-basic-function/ai-character-chat/how-to-create-ai-character/faqs-about-character-creation.md). # FAQs About Character Creation Here are some common questions you may encounter when creating a character. #### 1. How Do Tokens Affect Character Memory and Chat Consistency? Large language models (LLMs) process text in units called tokens. A token can be a whole word, part of a word, or a punctuation mark. For example, "Hello, world!" may use 4–5 tokens, depending on the model's tokenizer. Permanent tokens are the tokens used by character information that is included in each reply, primarily the Character Setting, along with the Scenario and Example Messages in most setups. The model's context window, which acts as its short-term memory, contains this permanent information plus the ongoing conversation history. Every chat model has a context window limit. The character settings, scenario, example messages, and chat history must all fit within this limit. When the limit is reached, older messages may be truncated, causing the model to lose track of earlier details and become less consistent. Free-tier models may have smaller effective context windows, making it especially important to keep permanent token usage low. SeaArt displays character counts rather than token counts. Each letter, space, and punctuation mark counts as one character. Character count can serve as a rough indicator of token usage, although the exact conversion varies by language and tokenizer. As a general estimate, one token corresponds to around 3–4 English characters on average. Keeping character settings concise leaves more room for conversation history, helping the model maintain continuity, remember earlier events, and respond more consistently during long chats. #### 2. What Causes Character Drift, and How Can You Prevent It? Character drift occurs when an AI character's personality, behavior, or established details change unexpectedly during a conversation. This can make the character feel inconsistent or cause it to forget previously established facts. Drift may result from contradictory instructions in the character settings or from earlier conversation details falling outside the model's available context due to token limits. To reduce character drift, focus on two priorities: consistency and token efficiency. Keep the Character Setting clear, concise, and free of conflicting instructions, as it typically makes up the largest portion of the permanent character information. When using a Scenario, ensure it aligns with the Character Setting and adds only relevant context. These steps help preserve more room for conversation history and support more consistent responses over long chats. Other fields may also affect the model's behavior, depending on how the platform and model process them. #### 3. How Can You Keep an AI Character's Personality Consistent? Consistency means keeping a character recognizable and believable across dozens or even hundreds of messages. Vague descriptions or contradictory traits can cause the model to switch unpredictably between different behaviors. **Core Guidelines for Character Settings** * **Choose Compatible Traits:** Avoid conflicting traits, such as "extremely confident" and "very shy," unless the contrast is intentional and clearly explained. * **Use Clear, Positive Statements:** Define what the character should do rather than what they should avoid. For example, use "always speaks politely" instead of "never swears unless angry." * **Focus on 4–7 Key Personality Traits:** Too many traits can dilute the character's defining qualities and increase the risk of contradictions. * **Use Familiar Archetypes:** LLMs often recognize common archetypes such as tsundere, catgirl, werewolf, and kuudere. A familiar term may communicate the concept more efficiently than a lengthy explanation. Add details when needed, especially if the character intentionally subverts the archetype. * **Keep the Backstory Relevant:** Include only the background details that influence the character's current personality or behavior. Omit information that doesn't contribute to the character's portrayal. **Tips for Using a Scenario** If you add a Scenario, make sure it aligns with the Character Setting in terms of personality, tone, and facts. Use it to establish the initial situation and context rather than repeating information already covered in the character definition. #### 4. How Can You Optimize Character Settings for Better Memory and Consistency? Efficiency means minimizing the tokens and characters used in permanent character information, leaving more room for conversation history within the model's context window. Lower permanent token usage can help reduce character drift during long conversations. **Core Guidelines for Character Settings** * **Remove Unnecessary Words:** Cut filler phrases and vague descriptions that add little useful information, such as "in her own way." * **Rely on the Model's Existing Knowledge:** Avoid explaining concepts the model already understands. For example, a description of a catgirl usually doesn't need to explain "cat ears and a tail" unless the character differs from the usual archetype. * **Use Precise Descriptors:** Choose one strong adjective instead of several overlapping ones. For example, "cold demeanor" is more concise than "cold, distant, and aloof." * **Keep the Backstory Focused:** Include only the essential events and motivations that explain the character's current traits and behavior. * **Avoid Repetition:** State each fact once instead of repeating the same information in different sections. **Tips for Using a Scenario** Keep the Scenario brief and use it only when the First Message alone cannot establish the necessary setting or context. Avoid repeating details already covered in the Character Setting. **Suggested Character Setting Lengths** The following ranges are rough guidelines rather than strict limits. The ideal length depends on the model, context window, and complexity of the character. * **Ideal:** 600–1,000 tokens (approximately 2,000–4,000 English characters). * **Generally acceptable:** Up to 1,200–1,500 tokens (approximately 5,000 English characters). * **Longer settings:** Above 1,500 tokens, character drift may become more likely when context space is limited, particularly during extended conversations. Applying these principles to the Character Setting and Scenario can help create characters that remain more consistent over long conversations without requiring advanced techniques. 🌟 Some of the information in this article is based on the guide [AI Character Creation - A Beginner's Guide \[Eng\]](https://www.seaart.ai/articleDetail/d6ii5nle878c73d4cmmg) by SeaArt creator - [Tearlament](https://www.seaart.ai/user/87ca7d040871c29ad4585fd2354388a6).