How does Janitor AI work?
Janitor AI pairs a character's written profile with a large language model. When you send a message, the model reads that profile plus your recent chat history and predicts an in-character reply. As of 2026-07 the default engine is the free JanitorLLM, and you can swap in outside models via proxy.
Why — the first-principles explanation
The engine underneath is a large language model (LLM), an AI trained to predict the next word given everything before it. By itself an LLM is a blank actor with no identity. Janitor AI gives it one by feeding it a character card: a hidden block of text describing the character's name, personality, backstory, and example lines. The model treats that card as standing instructions and answers as that character.
Every time you hit send, the system bundles three things and passes them to the model: the character card, a chunk of your recent conversation, and your new message. This bundle is the context window. Because it's limited, roughly 9,000 tokens on the free tier, only the most recent history fits, which is why bots eventually "forget" early details. Paying for Janitor+ enlarges that window about 5x.
The last piece is where the model runs. By default it's JanitorLLM on Janitor AI's servers. If you attach an API key and proxy, Janitor AI instead ships that same bundle to an outside model like OpenAI, which generates the reply and sends it back. The character logic stays the same; only the brain changes.
An example that makes it click
Imagine an improv actor wearing an earpiece. Before each scene, a director whispers a character sheet: 'You're a grumpy space pirate who loves puns.' Then, every time you speak, the director quickly reminds the actor of the last few lines and your newest line, and the actor responds in character. The earpiece can only hold so many reminders at once, so after a long scene the actor forgets how the story began. That earpiece is the context window.
Key facts
- Janitor AI runs on a large language model that predicts the next words in a conversation.
- Each character is defined by a 'character card' fed to the model as hidden instructions.
- The context window bundles the card, recent history, and your message; free tier is about 9K tokens.
- Bots 'forget' older messages once the conversation exceeds the context window.
- The default engine is JanitorLLM; external models can be attached via API key and proxy.
▶ The 60-second explainer (script)
How does Janitor AI actually work? Let's open the hood, as of July 2026. At the core is a large language model, an AI trained to predict the next word based on everything before it. On its own, it has no personality. Janitor AI gives it one using a 'character card', a hidden description of who the character is, their backstory, and how they talk. When you send a message, the system bundles three things together: that character card, a slice of your recent conversation, and your new message. It hands all of that to the model, which predicts a reply in character. That bundle has a size limit called the context window, about nine thousand tokens on the free tier, which is why bots eventually forget how a long chat started. Janitor Plus makes that window about five times bigger. And here's the flexible part: by default the brain is JanitorLLM, but if you attach an API key, Janitor AI sends the same bundle to an outside model like OpenAI instead. The character stays the same; only the brain changes.
What authoritative sources say
People also ask
Why do bots forget things I said earlier?
Because the context window is limited (about 9K tokens free), older messages drop out of the model's memory.
What is a character card?
A hidden block of text describing a character's name, personality, and style that the model follows as instructions.
Does Janitor AI train on my chats?
Chats are stored to provide memory and service; treat them as recorded, and avoid sharing sensitive personal details.
What changes when I use an external model?
Only the 'brain' changes. The character card and history are still sent, but a provider like OpenAI generates the reply.