What a context window is
A context window is the maximum amount of text an AI model can take in for a single reply. Everything the model knows about your story in that moment, from its instructions to the last line you wrote, has to fit inside it, and so does the reply it writes. Anything outside the window doesn’t exist as far as the model is concerned.
Windows are measured in tokens, the pieces models read and write. A token is often a whole common word and sometimes part of a longer one; in English, a token averages about four characters, or roughly three quarters of a word. Names, unusual spellings, and many other languages take more tokens per word.
So a 6,400-token window holds roughly 4,800 words in all, shared by everything in the prompt, and a 100,000-token window holds roughly 75,000.
Tokens in practice
Token counts are easier to feel with examples. At three quarters of a word per token:
| Text | Roughly |
|---|---|
| “I draw my sword and step into the rain.” | 12 tokens |
| A 150-word reply | 200 tokens |
| A 300-word lore card | 400 tokens |
| A 1,000-word premise | 1,300 tokens |
| A 100,000-word novel | 133,000 tokens |
The last row is the point. Even the largest plan’s 100,000 tokens can’t hold a novel-length story word for word, and a roleplay grows faster than it feels: a hundred exchanges of 150-word replies and 50-word turns already come to about 20,000 words, several times what the free plan’s window holds. That’s why every serious story app pairs its window with a memory system.
Counts vary by model, because each family of models splits text its own way. Rare names, invented words, and languages other than English usually take more tokens per word, so a world full of unusual names spends its window a little faster.
What fills the window on every reply
On LoreWeaver, each reply’s window is shared by:
- Instructions that tell the storyteller how to write and what never to decide for you.
- The world settings, which every reply reads: premise, tone, and style rules.
- Your character from the Author drawer, and your session notes, up to 500 tokens.
- Lore cards that are pinned or whose names came up.
- The recap of the session so far.
- Memories that match the moment.
- Recent turns, word for word, as many as fit.
- Loom directions, when a thread fires.
- Room for the reply, set by the reply length in generation settings.
Each part has a share. Recent story gets the largest share on every plan, and on paid plans lore comes next. Space that one part leaves unused flows on, first to the world’s settings and then to extra lore cards, and when many cards qualify, lore can borrow room from the story history. How LoreWeaver works walks through the same parts with a scene in progress, and lorebooks explained covers how cards compete for their share.
Context by plan
| Plan | Context allowance | Roughly, in words |
|---|---|---|
| Free | 6,400 tokens | 4,800 |
| Plus | 16,000 tokens | 12,000 |
| Pro | 32,000 tokens | 24,000 |
| Ultra | 64,000 tokens | 48,000 |
| Legendary | 100,000 tokens | 75,000 |
These are the most each plan allows. Some storytellers have smaller windows of their own, and LoreWeaver uses whichever limit is smaller. On the free plan, a lean world leaves most of the allowance for the story; Legendary’s allowance has room for tens of thousands of words of recent story, beside plenty of lore and memories. The plans page lists everything else each plan includes.
What happens when a story outgrows the window
Every long story does. When the recent turns no longer fit, the oldest ones leave the window first, and the one at the edge may be trimmed to fill the space exactly. From then on, older events reach the storyteller only through the other layers: memories that match the moment, the recap that covers the whole session, and lore cards that come up by name.
That hand-off is where details get lost if the layers are weak, and on LoreWeaver you can watch it. The context inspector under the composer shows the tokens used against the budget, how they split between instructions, world, lore, memories, recent turns, and your character, and which cards and memories were included, with the words that triggered each card. How AI memory works explains the layers in detail.
Is a bigger window always better?
Mostly, yes: more room means more of the story word for word, with more lore and memories beside it. But size isn’t everything:
- Speed and cost rise with every token sent.
- Attention is uneven. Research has found that models use information at the start and end of a long prompt more reliably than material in the middle (Lost in the Middle).
- Irrelevant text distracts. Stuffing the window with every card you own can make replies worse, not better.
That’s why selection matters as much as size. A focused prompt, with the right cards, the right memories, and a good recap, often beats a bigger one filled indiscriminately.
Context and memory do different jobs
It’s easy to treat context and memory as the same thing. They aren’t. Context is what the storyteller reads for this one reply. Memory is what LoreWeaver keeps between replies and chooses to put back into the context: memories, the recap, and lore cards.
A big window with no memory gives you a long, sharp present and no past: once a scene scrolls out, it’s gone. A small window with good memory gives you a shorter present and a past that returns when it matters. The best long stories have both, which is why paid plans widen the window and keep more memories per story at the same time.
In practice, a larger window helps most in dense scenes, such as a battle with many moving parts or a long negotiation, where the exact wording of recent turns matters. Memory matters most for everything older: the promise in the first session, the scar from the ambush, the name of the innkeeper’s daughter. How AI memory works follows that second half in detail.
Making every token count
- Keep the always-present parts lean. The premise, style rules, persona, and session notes are read on every reply, so trim them to what scenes actually need.
- Write short, specific lore cards. Split a card that covers two subjects, and put names and nicknames in its tags.
- Pin sparingly. Pinned cards take space on every reply.
- Watch reply length. Longer replies reserve more of the window for writing, which leaves less for the story. Cinematic, the preset for long replies, trades history for longer replies.
- Keep session notes current. They’re capped at 500 tokens, so replace old notes rather than piling on new ones.
- Check the inspector when a reply forgets something, before rewriting anything.
- Move up a plan when a story needs more of its past in view.
