What is a context window in AI chat?
Updated
A context window is the maximum amount of text a language model can take in at once, measured in tokens. It has to hold everything for a reply: the character card, instructions, your chat history and the reply itself. When a chat grows past it, the oldest messages are dropped, which is why AI characters forget early events. UnsavedAI uses about 61,000 tokens.
Try it now, free, no account neededStart a story with Silas CraneTired detective in a city that never driesWhat fills the window
Every time the model replies, the app sends it one big block of text. In a roleplay that block usually contains:
- the system prompt and any app instructions
- the character card: description, personality, scenario, example dialogue
- your persona, if you set one
- lorebook entries that fired
- as much of the chat history as still fits
- room for the reply being written
The model has no memory beyond this block. If a detail from message 12 is not in the window when message 400 is written, the model cannot know it, no matter how important it was.
A worked example
UnsavedAI models allow 61,440 tokens including the reply. A 1,500-token character card plus a 500-token persona leaves roughly 59,000 tokens. At around 200 tokens per message, that is a few hundred messages of history in view at once, a long evening of roleplay. Once the chat goes past that, the app trims from the top and keeps the newest turns.
In SillyTavern or JanitorAI you set a context size yourself. With UnsavedAI models, set it to 56,000 or less so there is always room for the reply. Setting it higher than the model allows makes long chats fail once they fill up rather than trimming quietly. The context size guide goes deeper.
Common mistakes
- Thinking a bigger window means perfect memory. Models pay most attention to the start and the most recent messages; details in the middle of a huge chat can still get fuzzy.
- Bloated cards. A 4,000-token description is read on every reply and pushes history out sooner.
- Setting max reply length to thousands of tokens. That space is reserved for the reply and taken from history.
- Never summarising. For very long stories, a short recap of key events keeps them in view after the original messages scroll out.
Context in UnsavedAI and other apps
Both UnsavedAI lanes, Gemma 4 12B on free and Gemma 4 31B on Priority, use the same 61,440-token window, so a long story fits on either. Character.AI and other apps do not always publish their context sizes, which is one reason memory feels so different from app to app. To test any of them, mention a small detail early in a story with Silas Crane, a detective who should notice it, and see whether he brings it up much later. For more memory tips, see making the AI remember your story.
Questions people ask
Is context the same as memory?
For most apps, yes: what the model can see in its context window is all it remembers. Some apps add memory features that summarise older events back into the window.
How many words is 61,000 tokens?
Roughly 45,000 English words, using the common rule of thumb of about three quarters of a word per token. The exact number depends on the text.
Why do replies slow down in long chats?
More context means more text to read before replying. Apps that keep the start of the prompt unchanged let the model reuse earlier work, which keeps long chats quicker.
