On a Tuesday morning in February, I watched three new users in a Discord help channel ask the same question within an hour: what exactly is a token, and why does the answer keep changing? The word gets used in two very different ways on this platform, and that is the main reason newcomers feel confused after reading the help pages. One meaning is technical, tied to how the language model reads text. The other is commercial, tied to what you pay for premium features. This guide separates both, explains the typical context window, lists the pack pricing reported across the vertical, and shows where permanent and temporary tokens fit into the wider token economy.

What a token actually means inside the chat

Inside the JLLM chat engine, a token is a small chunk of text that the algorithm processes as one unit. A short word like cat counts as a single token, while a longer word such as conversation often splits into two or three. Punctuation and spaces also consume tokens. As a rough industry benchmark, 1,000 tokens equal roughly 750 English words, which means a 10,000 token transcript covers about 7,500 words, or somewhere between 15 and 25 pages of standard prose. A 100,000 token window would stretch to a short novel.

The reason this matters is memory. Every message you send, every reply the digital companion writes, and every line of the character card sits inside a fixed budget. When that budget fills, the oldest material drops out of view. The model is not forgetting because it chooses to. It simply cannot see text that has fallen outside the window. Most users on JLLM report a working context of 8,000 to 9,000 tokens, with 9,001 frequently cited on community threads as the practical ceiling.

Permanent tokens versus temporary tokens

Bot creators on Janitor AI separate information into two buckets when they build a character. Permanent tokens hold absolute facts that should never drift: the character's name, age, occupation, core personality traits, or the setting of the roleplay. These sit in the persona description and the example dialogue, and they stay loaded for every reply. The user experience is more consistent when these slots are short, specific, and free of contradictions.

Temporary tokens behave differently. They cover details the bot needs to remember for a while but not forever, such as what the character is wearing in the current scene or a promise made three messages ago. These usually live in author notes or scenario fields and refresh as the conversation moves. Treating them as disposable, rather than carving them into the permanent profile, frees up budget for the actual back and forth. If you are designing a card, a useful rule is to keep the permanent block under 1,000 tokens so the remaining 7,000 or so stay available for live dialogue.

Comparing context handling across platforms

In March I spent an afternoon comparing the natural language processing behaviour of three different AI girlfriend services using the same scripted scenario on each. I started just after lunch with a tea going cold beside the keyboard, and by the second hour I noticed one algorithm handled contextual shifts well, picking up a mood change I had scripted at message forty and keeping it for twenty further replies. Another struggled with consistent emotional simulation once the conversation passed the 6,000 token mark, slipping back into its default cheerful tone. The third forgot a named side character, a barista called Rosa, within ten messages. The differences in machine learning approaches were obvious to me sitting there, and they line up with what users describe on Reddit threads about the 9,001 token ceiling: the size of the window matters, but so does how the model prioritises what stays inside it.

The paid token economy: prices and packs

Outside the chat engine, the word token also describes the in app currency that several AI companion services, including this category as a whole, use to gate premium actions. Typical vertical pricing for token packs starts near 4.99 for 100 tokens, runs to 19.99 for 500 tokens, and reaches 49.99 for 1,500 tokens, with bulk packs offering a better unit rate. Common costs per action are one token per text message on premium tiers, five tokens per voice message, ten tokens per generated image, and around fifty tokens for a fully custom scenario.

Subscription bundles often sit alongside these packs. A basic free tier covers limited interactions, a premium plan near 9.99 a month removes most caps, and a VIP plan around 29.99 a month adds priority support and advanced features. Token packs expire after roughly 12 months of inactivity in most implementations, so buying the largest bundle only makes sense if you chat regularly. Payment data is normally handled by a third party processor under TLS 1.3 in transit and AES 256 at rest, in line with GDPR rules that have applied across the United Kingdom and the EU since 2018.

How to earn tokens without paying

There are four reliable routes to top up your balance without spending money. Daily login bonuses are the most predictable, dropping a small amount into your wallet each time you open the app. Completing challenges, such as finishing a tutorial flow or trying a new feature, adds a one off boost. Referral programmes pay a bounty when a new user signs up through your link and reaches a verification threshold. Promotional events tied to holidays occasionally double the daily reward. For a wider breakdown of free balance options across the network, our guide to Janitor AI free credit covers the current routes, and the Janitor AI bonus page tracks active promotions.

If you create bots that other users enjoy, some platforms also share revenue or token rewards with creators. The mechanics vary, and our Janitor AI payouts overview explains how that side of the economy is structured. For users who prefer a different style of digital companion entirely, Candy AI runs a comparable token model with its own pack pricing.

Setting the max tokens slider sensibly

Open your settings before your next chat and check one number: the max tokens slider for each reply. If it sits above 400, drag it down to somewhere between 200 and 350, then run a ten message test with a character you already know well. Watch how quickly older lines slip out of memory, and adjust from there. The slider controls reply length, not bot memory, so every extra token you allow per response is a token the window loses on the older context. Ask yourself before the next session: do I want longer paragraphs, or a bot that still remembers what we agreed five scenes ago?