Training

Dyva characters aren't static. Every conversation is a learning opportunity. When users chat with your character, the system observes patterns — which responses land well, which ones fall flat, and where the character drifts off-personality. Over time, these signals compound to make your character sharper and more consistent.

This learning happens at two levels. At the system level, aggregated feedback across all conversations helps the underlying model understand what good responses look like for your character. At the individual level, the memory system stores facts, preferences, and relationship context for each user, so the character can personalize future interactions.

You don't need to do anything special to activate learning — it happens automatically. But you can accelerate it by actively testing your character, reviewing conversation logs, and providing explicit feedback. Characters that are actively maintained by their creators improve much faster than characters that are published and forgotten.

Memory is one of Dyva's most powerful features. When a user tells your character their name, their favorite color, or that they're studying for a physics exam next Tuesday, the character can remember all of that — not just for the current chat, but across sessions.

The memory system works in layers:

• Short-term memory — The current conversation context. The character remembers everything said in the active chat session, including the tone and direction of the conversation.
• Long-term memory — Key facts, preferences, and events extracted from past conversations. These are stored persistently and retrieved when relevant. If a user mentioned their dog's name three weeks ago, the character can recall it naturally.
• Relationship memory — The character tracks its relationship with each user: how often they chat, what topics they gravitate toward, and the emotional tone of their interactions. This allows the character to adjust its behavior — being more casual with a regular, more welcoming with a new visitor.

Memory extraction is automatic. After each conversation, the system identifies important facts and stores them. As a creator, you can view what your character has memorized in the dashboard and manually correct or remove memories that are inaccurate or undesirable.

Users can rate individual messages from your character with a thumbs up or thumbs down. This feedback is one of the most valuable signals available to you as a creator. It tells you exactly which responses resonated and which ones missed the mark.

You can view feedback in your creator dashboard. Look for patterns: if multiple users thumbs-down a certain type of response (e.g., the character breaks character when asked about a specific topic), that's a clear signal to update your definition to address that scenario.

Beyond individual ratings, your character receives an overall quality score based on aggregated feedback, conversation length (longer conversations suggest engagement), and return rate (users coming back to chat again). This score influences your character's visibility in the marketplace — higher-quality characters get more prominent placement.

Don't be discouraged by negative feedback. Every thumbs-down is a free lesson in what to fix. The best creators treat feedback as a compass: it points directly at the parts of their character that need work.

Here are proven strategies for making your character better over time:

• Chat with your own character regularly. You'll catch issues that don't show up in the definition alone. Try to break it — ask edge-case questions, push emotional boundaries, see what happens when you contradict it.
• Read your conversation logs. The dashboard shows recent conversations (anonymized). Skim through them to spot recurring issues: does the character repeat certain phrases? Does it lose its personality after long conversations? Does it handle specific topics poorly?
• Iterate on the definition. After each round of testing, make small, targeted changes. Don't rewrite the entire definition at once — tweak one thing, test it, and see if it helps. Small changes are easier to evaluate.
• Use the greeting as a calibration tool. If users frequently misunderstand what the character is about, your greeting may not be doing its job. A stronger greeting sets better expectations and leads to better conversations overall.
• Benchmark against the best. Chat with popular characters in your category. Notice what they do well. You're not trying to copy them — you're studying what "good" looks like so you can aim higher.

Quality is a moving target. As the platform evolves and user expectations grow, the characters that stay relevant are the ones whose creators keep showing up and making them better.