Context Window

You might say

The AI forgets what I said at the start of a long chat. Does it have a memory limit?

The total amount of input and output a model can process in one requestAfter a long chat, “follow the first requirement” can fail if that requirement was not included in the current request. More context capacity does not make every piece of information relevant.

Know first

Token

Context Length

What fills the space for one answer?

This request allows up to 1,000 illustrative Tokens560 input + 200 output reserved

System rules180 Token (illustrative)

Current task120 Token (illustrative)

login.js260 Token (illustrative)

It still fits, with room reserved for the model’s answer.

How context window space is allocated

Shared space: System prompts, conversation history, files, and response space all share the same total context window limit.

Output truncation is separate: Input context and max output tokens are separate (output truncation is usually due to reaching the output limit, not necessarily a full window).

Bigger isn't always better: A larger window doesn't mean you can cram everything in—irrelevant material dilutes the AI's attention and causes it to drift.

You can say this to an AI Agent

Before continuing this task, list the rules, files, and recent decisions that must remain, then compress older material into a short working brief. Show me that brief and the material you recommend excluding for confirmation before proceeding; do not assume you can delete the current conversation.

Learn next

1Token→2Context Engineering→3Token Usage & Cost→

Recommended tool

Web fine-tuning tool · Selector Pro

Select page elements for precise AI edits

Click a button, text, or card, note the change, then copy its exact location to AI.

Open Selector Pro website

An Agent starts missing newly added requirements and replies often cut off. Before continuing the same task, what should you do first?

A. Send another message and paste the remaining material thereB. Identify the rules, material, and recent decisions the task needs, summarize older content, and reserve answer spaceC. Keep every input and reduce the response reserve as much as possible

ANS Original Methods

Apply this term in a real workflow

Multi-Agent & Project Governance

Multi-Agent Team Mode, 1:20:100 Cost Routing & Long-Project Governance

Replace single-model brute force with Explorer, Executor, Reviewer, and Advisor sub-agents, pairing pre-coding Explore and pre-commit Simplify with a lean AGENTS.md project index.