Arcanum

The AI Roleplay Token Budget: What Your System Prompt, World File, and Memory Block Actually Cost

· by Rukka

Every AI roleplay session runs on a fixed amount of room, and almost everyone playing one has no idea how much of it they’ve already spent before the first turn even happens. A token budget is simply the accounting of that room: how many tokens your system prompt, world file, character cards, and memory block cost before the conversation starts, and how much is left over for the part that actually matters — playing the game.

Most players and even most platform builders never do this math. They write a prompt, load a world file, and find out the hard way — a session that goes stale by turn thirty on one platform and turn three hundred on another, using what feels like the same amount of setup. It isn’t the same amount. Here’s what actually costs what, and how to stop guessing.

What Actually Eats Your Context Window

Four things typically compete for space before a single turn of play happens, and they’re not equal:

System prompt. Your genre, tone, agency rules, and continuity instructions. This should be the cheapest fixed cost in the budget — it’s written once and never changes size.

World file, lorebook, or setting bible. Faction names, locations, history, house rules. This is where budgets quietly balloon, because “just a little more lore” feels free until you total it up.

Character cards. Personality, scenario, opening message, and example dialogue for every character the AI is voicing. One card is cheap; a party of six companions each with a full card is not.

Memory block or campaign summary. Whatever you’re carrying forward from previous sessions — a running state log, a compressed summary, or (on some platforms) an external database that doesn’t touch the context window at all.

Everything in that list is a fixed cost — it’s the same size on turn one as it is on turn three hundred. The conversation itself, growing one exchange at a time, is the variable cost. Both draw from the same pool, and the fixed cost is the one entirely within your control before you’ve typed a single line of the actual story.

How Many Tokens Is That, Really?

Numbers help more than adjectives here. Using the rough industry rule of thumb — about 4 characters or 0.75 words per token — some worked examples:

  • A lean system prompt (genre, tone, agency rules, continuity instructions, no flavor text): roughly 150–300 tokens.
  • A bloated system prompt written as prose instead of instructions: 800–1,000+ tokens for the same amount of actual guidance.
  • A well-written character card in the 300–600 word range most card-writing guides recommend: roughly 400–800 tokens.
  • The same character over-written as a 2,000-word biography: 2,500–3,000+ tokens — three to four times the cost for information the model uses just as well in a third the space.
  • A flat world file pasted as one block — a setting bible, faction list, and history running 2,000–4,000 words: roughly 2,700–5,500 tokens, present in full on every single turn whether it’s relevant or not.
  • A compressed memory block, built to the under-500-word target the Campaign Memory Tool uses: capped around 650 tokens by design, no matter how long the campaign has actually run.

None of these numbers are exact — no provider publishes a client-side tokenizer for every model, so this is a consistent estimate, not a guarantee. But the relative sizes are the point: a lean setup and a bloated one can hold the exact same information at a three- or four-times difference in cost, and that difference is pure budget you’re either keeping or throwing away.

The Real Budget Killer Isn’t Your Fixed Content — It’s Which Model You’re On

Here’s the number that actually matters: on a flagship model with a 1,000,000-token context window, even a genuinely heavy fixed-content setup — a padded system prompt, a full flat world file, an uncompressed memory block, call it 6,000–8,000 tokens combined — is well under 1% of the total budget. On that model, budgeting your fixed content barely moves the needle. The conversation itself, turn after turn, is what eventually fills the window.

The picture changes completely on a smaller or free tier. AI Dungeon’s free Wanderer tier caps out around 4,000 tokens — its paid tiers step up to 8k, 16k, and 32k. A local model run through SillyTavern often sits in a similar range unless you’ve got the hardware for something bigger. On a 4,000-token budget, that same 6,000–8,000-token fixed setup doesn’t just eat a slice of the budget — it doesn’t fit at all. Trim it to a lean system prompt, a tight character card, and a lorebook (not a flat file), and the same 4,000-token budget can comfortably support real play.

This is why the question “how big is my fixed content?” only means something in relation to “what model am I actually running on.” The free/cheap tiers most people and most indie platforms actually build on are where token budgeting stops being an academic exercise and starts being the difference between a session that plays and one that errors out before it starts.

Lorebooks vs. Flat Files: The Biggest Lever You’re Not Using

If there’s one architectural choice that matters more than any individual number above, it’s this: a lorebook (the SillyTavern approach, also used in various forms elsewhere) only injects entries relevant to what’s currently being discussed, scanning recent messages for keyword matches and pulling in just the matching lore. A 6,000-token world file organized as a lorebook might contribute a few hundred tokens on any given turn — the desert kingdom’s entry doesn’t load while the party is in the northern forest.

Paste that same 6,000 tokens as one flat block in a system prompt or a Custom GPT’s knowledge field, and all of it sits in context on every single turn, relevant or not. Same information, wildly different cost. If your platform or setup supports keyword-triggered lore and you’re not using it, this is the single highest-leverage change available before you touch anything else.

Calculate Your Own Budget

Rather than estimating in your head, the free Context Budget Calculator does this math directly: pick your model — Claude, ChatGPT, Gemini, Grok, DeepSeek, or MiniMax, flagship or cheap tier — paste in your actual system prompt, world file, and memory block, and it shows exactly how many tokens each one costs, how much room is left for play, and roughly what turn you’ll hit the wall based on how many tokens an average exchange costs you. No account, no API calls — everything runs in your browser.

It’s the fastest way to answer the question this whole article is about: not “is my setup too big” in the abstract, but too big for the specific model you’re actually running it on.

How to Cut Your Budget When You’re Close to the Wall

If the calculator (or the math above) shows your fixed content eating more than it should, the fixes are the same regardless of platform:

  • Convert a flat world file into a lorebook. The single biggest win available, if your setup supports it — see above.
  • Trim the system prompt to instructions, not prose. Rules the model needs to follow, not paragraphs describing the vibe you’re going for.
  • Rewrite bloated character cards. Specificity beats length — a tight 300–600 words consistently outperforms a 2,000-word biography, at a third of the token cost.
  • Compress your memory instead of carrying the raw transcript. The Campaign Memory Tool turns a sprawling session history into a structured summary under 500 words, restoring most of your budget without losing the facts that matter.
  • Move deep lore off standing context entirely. Anything you’d only reference occasionally — deep backstory, a rarely-visited location — doesn’t need to sit in the budget every single turn if your platform can pull it in on demand instead.

For the mechanics behind why all of this matters in the first place — what happens once you actually run out of room — see why AI campaigns fall apart at turn 50.

Frequently Asked Questions

How many tokens does an AI roleplay system prompt need? A lean, effective system prompt runs roughly 150 to 300 tokens — genre, tone, agency rules, and continuity instructions, written as direct instructions rather than flavor text. Prompts that balloon past 800 to 1,000 tokens are usually padded with prose the model doesn’t need to follow the rule, and every extra token there is one less available for the actual conversation.

How many tokens is a character card? A well-written card in the 300 to 600 word range most style guides recommend works out to roughly 400 to 800 tokens once personality, scenario, an opening message, and example dialogue are included. A bloated 2,000-word biography can run 2,500 to 3,000 tokens for the same amount of usable information, which is pure budget waste.

Do lorebook entries cost the same tokens as a flat world file? No, and this is the single biggest lever available. A lorebook only injects entries relevant to what’s currently happening, so a 6,000-token world file might contribute a few hundred tokens on any given turn. Pasting the same content as one flat block puts all of it in context on every single message, whether it’s relevant or not.

Why does my AI roleplay run out of context faster on some platforms? Because context windows vary enormously by tier, not just by model. AI Dungeon’s free Wanderer tier caps around 4,000 tokens, where a flagship API model can run past 1,000,000 — a 250x difference. The same system prompt and world file that are a rounding error on one are a third of the entire budget on the other.

Does a bigger context window mean better AI roleplay? Not by itself. A large context window is a ceiling, not a guarantee of quality — most models’ recall degrades well before they hit their technical limit, a pattern often called context rot. A lean, well-budgeted setup on a mid-sized context window frequently holds together better than a bloated one that happens to fit inside a huge one.

How do I calculate my own AI roleplay token budget? Paste your system prompt, world file, and memory block into the free Context Budget Calculator, pick your model, and it shows exactly how many tokens each piece costs, how much room is left for actual play, and roughly what turn you’ll hit the wall based on your average turn length.