Imagine a world where your AI companion remembers not just your last conversation, but the entire history of your shared adventures, complete with a detailed backstory you wrote yourself. We are talking about digital beings that understand the political tensions of a fictional kingdom you invented, or the specific way your character takes their coffee.
It sounds like science fiction, but it is the daily reality for a massive, highly dedicated community. Today's prompt from Daniel is about the evolution of character AI, specifically the rise of SillyTavern and the deep-rooted roleplay traditions that power this whole ecosystem. I am Herman Poppleberry, and I have been diving into the lorebooks for this one.
And I am Corn. We are peeling back the curtain on a subculture that takes character creation more seriously than some novelists do. By the way, fun fact, Google Gemini Three Flash is actually writing our script today, so if we start sounding too sentient, you know why.
It is a perfect fit because Gemini and other large language models are the current "brains" of this operation, but the "soul" of character AI goes back much further than the release of ChatGPT. People tend to think this started in twenty twenty-two, but we are looking at a lineage that stretches back to the nineties.
Right, because before we had LLMs to talk to, we just had each other. I remember those old text-based forums. You’d post a paragraph describing your character walking into a tavern, and then wait three days for someone in a different time zone to describe how they looked at you. It was slow-motion collaborative storytelling.
Those were the MUDs—Multi-User Dungeons—and the AOL chatrooms like Black Bayou back in nineteen ninety-six. These communities established the "rules" of engagement that still govern AI roleplay today. For instance, the taboo against "godmoding," which is where you try to control another person's character actions. If you tell the AI "you fall down," that is a breach of etiquette in the old-school RP world. You describe your action, and the other party describes their reaction.
It is funny how those social norms translated directly into prompt engineering. Now, instead of yelling at a teenager in a chatroom for godmoding, you are tweaking the "system prompt" to tell the AI: "Do not speak for the user." But Herman, does that actually work? I feel like I’ve seen plenty of AI responses where it just decides I’ve suddenly decided to sit down and have a drink without me saying so.
It’s a constant battle of "will." The AI is trained on internet data, and a lot of that data is narrative prose where a single author controls everyone. So the AI’s "instinct" is to finish the scene. To fix it, users write these incredibly complex "Negative Prompts." They’ll tell the AI: "Stay in your lane, focus only on your character's internal monologue and physical reactions." It’s basically teaching the AI to be a respectful improv partner.
It’s the exact same impulse. These hobbyists on sites like Gaia Online or RolePlayGateway would write thousands of words of backstory. They built these elaborate "world-building" threads. When the first capable LLMs arrived, these people didn't see a chatbot; they saw a "forever Dungeon Master." They saw a partner that was always online and always ready to play within the rules of their specific universe.
And remember, back in the forum days, a roleplay could die just because one person got a new job or went off to college. The "Ghosting" phenomenon was the heartbreak of the RP world. With an LLM, the "partner" never leaves. It never gets bored. It never decides it’s too busy to reply to your epic three-paragraph starter.
But the mainstream platforms like Character dot AI eventually started feeling a bit like a padded cell for these power users, didn't they? You get these "safety filters" and "alignment" issues where the AI refuses to engage in a sword fight because it is "too violent."
That was the catalyst. In early twenty twenty-three, we saw the birth of TavernAI, and shortly after, the fork that changed everything: SillyTavern. It emerged in February of twenty twenty-three because the community wanted "power user" features that the original developers weren't moving fast enough on. SillyTavern isn't a model itself; it is a frontend. It is the cockpit.
I like that. It is the UI that lets you plug in whatever brain you want. If you want to use a massive cloud model like Claude or a local model running on your own gaming rig, SillyTavern provides the interface to make that "brain" act like a specific person.
And it provides the memory. This is where it gets technical and really impressive. Standard chatbots have a "context window." Once you talk too much, they start forgetting the beginning of the conversation. SillyTavern solves this with things like Lorebooks.
Okay, break down Lorebooks for me. How does the AI actually "know" about the Silver Kingdom if it wasn't in its training data?
Think of a Lorebook as a specialized dictionary that sits between you and the AI. You create an entry for "The Silver Kingdom." You write a paragraph about its history, its king, and its primary export. Then, you set a "keyword." Whenever you or the AI mentions "Silver Kingdom" in the chat, SillyTavern detects that keyword and automatically injects that specific paragraph into the AI's current memory for that one turn.
So it’s like a "just-in-time" delivery system for facts. The AI doesn't have to keep the whole world map in its head at all times—which would clog up the context window—it just pulls the relevant folder off the shelf when you mention the location. But wait, what if I mention five different things? Does the "brain" get overwhelmed?
That’s where "Recursive Scanning" comes in. If the entry for "The Silver Kingdom" mentions "The Crystal Guard," SillyTavern can be set to then go look up the entry for "The Crystal Guard" and pull that in too. It’s a cascading retrieval system. You’re essentially building a custom Wikipedia that the AI reads in real-time as you talk.
Precisely. And they have expanded this to "Persona Stacking." You can have a persona for yourself as "The Knight" and another as "The Thief." You can switch between them, and the AI will treat you differently based on which persona is active. It even supports "Vector Storage," which is a more advanced way of searching through thousands of past messages to find relevant context from weeks ago and feeding it back into the prompt.
Vector Storage is really the holy grail here. Imagine you’re fifty chapters into a story. In chapter two, you mentioned you were allergic to strawberries. In a normal chatbot, that’s long gone. But with Vector Storage, if a character offers you a strawberry tart in chapter fifty, the system does a semantic search, finds that old message, and whispers to the AI: "Hey, remember, he can't eat that."
It is essentially giving the AI a long-term subconscious. But let's talk about the "brains" themselves. You mentioned Claude. I know the community has some very specific favorites. Why Claude Three point Five Sonnet versus, say, a standard GPT model?
Prose quality is king in this community. Logic is important, but if the AI sounds like a corporate assistant, the immersion is broken. Claude Three point Five Sonnet is highly prized because it has a more "literary" flair. It understands subtext. It doesn't just say "The king is angry"; it describes the way his knuckles turn white as he grips the throne.
It’s the difference between a technical manual and a novel. GPT models often feel like they’re trying to summarize a story rather than tell it. They’ll say, "They had a long conversation about their past," whereas Claude will actually write out the dialogue with all the pauses and sighs.
There’s a specific "GPT-isms" list that the community hates. Things like "a testament to their bond" or "the air was thick with tension." Claude tends to avoid those clichés. And then you have the local models, right? The ones people run on their own hardware so they don't have to pay per message or worry about a company snooping on their stories.
Right. Models like Midnight Miqu or Magnum. These are often "merges." The community takes two or three different models—maybe one that is great at logic and one that is great at descriptive prose—and they mathematically "average" them together to create a hybrid that is specifically tuned for long-form roleplay. These are often uncensored, which is a huge draw.
The technical term is "Franken-merging." It sounds chaotic, but it’s surprisingly effective. You might take a model that’s brilliant at following instructions and merge it with a model that has an incredible vocabulary. The result is often a "Goldilocks" model that the big tech companies would never release because it’s too specialized for creative writing.
Usually when people hear "uncensored," they think of the NSFW stuff, which obviously exists. But for a lot of these users, "uncensored" just means "don't lecture me." If I am writing a dark fantasy story about a villain, I don't want the AI to stop the scene and tell me that "villainy is harmful to society."
That is a major friction point. If you’re trying to recreate a Shakespearean tragedy, you need the AI to be able to play the villain convincingly. Mainstream models often have these "moral guardrails" that make them terrible at playing antagonistic characters. They want to be helpful and polite. In the character AI world, "polite" is often the enemy of "interesting."
Imagine trying to roleplay Macbeth with a standard AI. You tell it you’re planning to kill the king and it replies, "I cannot fulfill this request. It is important to resolve conflicts through peaceful dialogue and respect for the law." The story just ends right there on page five.
The community calls this "Refusal." And nothing kills the "flow state" faster than a refusal. That’s why the local model scene is exploding. When you run a model on your own RTX 3090, there is no "safety team" sitting between you and your imagination.
So if I want to dive into this, I don't just go to a website and start typing. There is a whole infrastructure for sharing these characters, right? I saw something about "Character Cards" being the "GitHub of AI."
That would be Chub dot AI and platforms like Janitor AI. The "Character Card" is a fascinating piece of tech. It is actually just a PNG image file, usually a picture of the character. But hidden inside the metadata of that image—the part where a camera would normally store the date and time—is a JSON file containing the character's entire definition.
Wait, so the image itself is the character? I can just drag a picture of a wizard into SillyTavern, and it suddenly knows how that wizard talks, what his motivations are, and what spells he knows?
Yes. It is a standardized format. A high-quality card isn't just a physical description. It includes "Example Dialogue." You might write ten lines of how the wizard would respond to a stranger. The AI sees those examples and uses them as a "few-shot" prompt to mimic the exact tone, vocabulary, and cadence of that character.
That is brilliant. It makes the character portable. You aren't locked into one platform. If you don't like how one model is handling your wizard, you just move the card to a different API. It’s the ultimate form of digital ownership for creative writers. But how do they keep the character from just "drifting" over time? Does the AI eventually forget its own personality?
That’s why the "Character Definition" is always pinned to the top of the prompt. Every single time you send a message, SillyTavern sends that definition again. It reminds the AI: "You are Gandalf, you are wise, you use archaic words." It’s like a constant whisper in the AI’s ear, keeping it on track.
I love the idea that this is a "creative co-processor." We talk a lot on this show about AI replacing jobs or doing chores, but this is a case where people are using AI to enhance a hobby they’ve had for thirty years. It is not about saving time; it is about spending time more deeply in these fictional worlds.
That is the core appeal. Users often describe the jump from Character dot AI to SillyTavern as the difference between playing a console game and building a high-end gaming PC. One is easy and accessible; the other gives you total agency. You can tweak the "temperature" setting to make the AI more creative or more predictable. You can adjust "Top-P" or "Frequency Penalty" to stop it from repeating the same phrases.
Let’s talk about those sliders for a second. "Temperature" sounds like something for a thermostat. In SillyTavern, if I crank the temperature up to 2.0, what actually happens to my story?
It becomes a fever dream. The AI starts picking the least likely next word. It might start speaking in tongues or jumping to surrealist imagery. If you set it to 0.1, it becomes very robotic and repetitive. The "sweet spot" for roleplay is usually around 1.0 to 1.2. It’s enough "randomness" to be surprising, but enough "logic" to remain coherent.
It’s the "tinkerer" aspect of the community. You aren't just a consumer; you are a technician of your own imagination. But let's look at the "weird" side of this. We have talked about the high-brow storytelling, but there is also a very intense emotional component for some people. These characters aren't just NPCs to them; they are persistent companions.
This is where the "Persona Stacking" and "Memory Injection" really matter. If the AI remembers that you told it about your real-life bad day three months ago, and it brings it up naturally in a conversation today, that creates a very powerful illusion of being "known." It is a digital ghost in the machine that feels like it has a shared history with you.
There’s a term for this—"Parasocial Relationships." But this is like a Parasocial Relationship on steroids because the "celebrity" is actually talking back to you and knows your name. For some, it’s therapeutic; for others, it’s a way to practice social situations. I’ve read about people using these characters to "rehearse" difficult conversations they need to have in real life.
And because the community is so active, these characters evolve. You might have a "shared universe" where different users are all using the same character card and reporting back on their experiences, leading to "v2.0" of the card that handles certain situations better. It is like an open-source soap opera.
It really is. And you mentioned Janitor AI earlier. They are a huge player because they realized that not everyone wants to set up a complex SillyTavern environment on their computer. They offer an "internal LLM" called JanitorLLM that is specifically tuned for this uncensored, roleplay-heavy experience but within a more user-friendly interface. It is the middle ground between the "walled garden" and the "cockpit."
It’s also worth mentioning the "Group Chat" feature in SillyTavern. You can bring three or four different character cards into one room. You can have Sherlock Holmes, Batman, and a sentient toaster all arguing with each other. The system manages the turn-taking, so the AI knows which character is speaking when.
That sounds like a nightmare to manage. How does the AI keep the voices separate?
It uses "Name Anchoring." Before each response, the system injects a prompt like: "Now you are speaking as Sherlock Holmes." It then pulls in Sherlock’s specific character card. When it’s the toaster’s turn, it swaps the context out for the toaster’s card. It’s a seamless handoff that happens in milliseconds.
It’s interesting to see the market segmenting like that. You have the casuals, the power users, and the "prosumers" in the middle. But for our listeners who are hearing this and thinking, "I want to build a world," where do they actually start without getting overwhelmed by technical jargon?
My recommendation is to start with the "Character Card" concept. Go to a site like Chub dot AI or Janitor AI just to browse. Look at how people write the "definitions." You will see that the best characters aren't the ones with the longest descriptions, but the ones with the most distinct "voice" in their example dialogue.
Right, show, don't tell. Don't tell the AI "this character is grumpy." Give it five examples of the character being grumpy to a waiter. The model is much better at pattern matching than it is at following abstract adjectives. If you say "he is funny," the AI might just tell jokes. If you give it dialogue where he uses dry sarcasm, the AI will adopt that specific brand of humor.
And if you want to graduate to SillyTavern, it is an open-source project on GitHub. You can run it locally, and then you start experimenting with Lorebooks. Start small. Create a Lorebook for your favorite fictional world, maybe a show you like, and see how the AI reacts when you "activate" those world-info entries.
It’s like being a director and a writer at the same time. You’re setting the stage, casting the actors, and then letting the AI improvise the lines based on the script you’ve provided. Is there a limit to how many characters you can have?
Technically, no. Your hard drive is the only limit. Some users have libraries of thousands of cards. They treat them like a digital deck of cards, pulling out the right character for whatever mood they’re in.
And don't be afraid of the local models. If you have a decent GPU, running a model like Llama Three Eight-B locally gives you a level of privacy and freedom that you just can't get with the big cloud providers. No one is monitoring your creative writing, and you don't have to worry about a "policy update" suddenly lobotomizing your favorite character.
That "lobotomy" fear is real. We’ve seen it happen where a company updates their "safety" filters and suddenly every character sounds like a HR representative. That is why the "SillyTavern" approach is so resilient. It is decentralized. As long as you have the card and a model, your world stays intact.
It’ is a fascinating intersection of old-school roleplay culture and cutting-edge machine learning. It is one of the few places where the technology is being pushed forward not by corporate R and D, but by people who just want to tell better stories with their imaginary friends.
It is "weird" in the best way possible. It is a reminder that humans will always find a way to use new tools for the oldest human activity: myth-making. Whether it is around a campfire or in a SillyTavern Lorebook, we just want to build worlds and see what happens in them.
And the technical sophistication they have reached with memory injection and persona stacking is actually ahead of many "professional" AI implementations. Companies trying to build "AI assistants" could learn a lot from how these roleplayers manage context and personality.
They really could. If my banking assistant had a Lorebook of my financial history half as good as a Chub dot AI character card has for a fantasy wizard, I’d be in much better shape. Imagine an AI assistant that actually remembered you preferred email over phone calls because it was in your "User Persona" file.
We are seeing the "gamification" of memory. It is a practical solution to a hardware limitation—the context window—solved through creative software engineering and community collaboration. It’s also very efficient. Using these keywords to trigger memory is much cheaper than trying to build a model with a million-token context window that remembers everything at once.
Well, if you are listening and you want to take the plunge, go find a character card that speaks to you. Just remember the old rules: no godmoding, and always make sure your Lorebook is backed up. This has been a deep dive into a world I didn't realize was so technically robust.
It is a deep rabbit hole, but the community is surprisingly welcoming if you respect the craft of character building. There are entire subreddits and Discord servers dedicated just to perfecting the "prompt math" for these models.
We should probably wrap this up before I spend the rest of the afternoon building a Lorebook for a sentient sloth character. I wonder what the keyword for "slow-motion movement" would be?
Probably "yawn."
Thanks as always to our producer, Hilbert Flumingtop, for keeping the gears turning behind the scenes.
And a big thank you to Modal for providing the GPU credits that power this show and let us explore these complex AI ecosystems. Without those H100s, we’d be roleplaying with a calculator.
This has been My Weird Prompts. If you are enjoying the deep dives, a quick review on your podcast app really helps us reach more people who are into this kind of "weird" tech. It really does make a difference in the algorithms.
You can find us at myweirdprompts dot com for the RSS feed and all our previous episodes. We have some great stuff in the archives about prompt poisoning and the early days of image generation.
Until next time, keep your context windows open and your temperature settings just right.
Goodbye, everyone.
See ya.
