Context Management

ℹ️Info

Quick Reference

• Context = All information Caret knows about your project
• Context Window = Maximum information Caret can process at once (varies by model)
• Token = Unit of text measurement (~3/4 of an English word)
• Auto-management = Caret automatically handles context through Focus Chain & Auto Compact

What is Context Management?

Context management is how Caret maintains understanding of your project throughout a conversation. Think of it as the shared memory between you and Caret - containing code, decisions, requirements, and progress.

The Three Layers of Context

1. Immediate Context - Current conversation and active files
2. Project Context - Your codebase, structure, and patterns
3. Persistent Context - Memory Bank, .clinerules, and documentation

Understanding Context Windows

Every AI model has a context window - the maximum amount of information it can process in a single conversation. This is measured in tokens:

Token Limits by Model

Model	Context Window	Effective Limit*	Best For
Claude 3.5 Sonnet	200,000 tokens	150,000 tokens	Complex tasks, large codebases
Claude 3.5 Haiku	200,000 tokens	150,000 tokens	Faster responses, simpler tasks
GPT-4o	128,000 tokens	100,000 tokens	General purpose development
Gemini 2.0 Flash	1,000,000+ tokens	400,000 tokens	Very large contexts
DeepSeek v3	64,000 tokens	50,000 tokens	Cost-effective coding
Qwen 2.5 Coder	128,000 tokens	100,000 tokens	Specialized coding tasks

*Effective limit is ~75-80% of maximum for optimal performance

💡Tip

Token Math Made Simple

• 1 token ≈ 3/4 of an English word
• 100 tokens ≈ 75 words ≈ 3-5 lines of code
• 10,000 tokens ≈ 7,500 words ≈ ~15 pages of text
• A typical source file: 500-2,000 tokens

How Caret Builds Context

Building effective context is what makes Caret truly useful. When you start a task, Caret doesn't just passively wait for information - he actively gathers context about your project, asks clarifying questions when needed, and adapts to what's happening in real-time. This combination of automatic discovery, user guidance, and dynamic adaptation ensures Caret always has the right information to solve your problems effectively.

1. Automatic Context Gathering

When you start a task, Caret proactively:

What Caret automatically discovers:

• Project structure and file organization
• Import relationships and dependencies
• Code patterns and conventions
• Configuration files and settings
• Recent changes and git history (when using @git)

2. User-Guided Context

While automatic discovery handles much of the work, you control what Caret focuses on. The more specific and relevant context you provide, the better Caret can understand your needs and deliver accurate solutions.

You enhance context by:

• @ Mentioning files, folders, or URLs
• Providing requirements in natural language
• Sharing screenshots for UI context
• Adding documentation through .clinerules or Memory Bank
• Answering questions when Caret needs clarification

3. Dynamic Context Adaptation

Caret adapts context dynamically throughout your conversation. It considers the complexity of your request, available context window space, current task progress, error messages and feedback, plus previous decisions made during the conversation to determine what information matters most at each step.

The Context Window Progress Bar

Monitor your context usage in real-time:

Understanding the Indicators

• ⬆️ Input Tokens: Information sent to the model (your messages + context)
• ⬇️ Output Tokens: Model's responses and generated code
• ➡️ Cache Tokens: Previously processed tokens that's reused (reduces costs and improves speed)
• Progress Bar: Visual representation of usage
• Percentage: Current usage of total capacity

Automatic Context Management Features

Caret includes intelligent systems that handle context for you:

Focus Chain (Default: ON)

Focus Chain maintains task continuity through automatic todo lists. When you start a task, Caret generates actionable steps and updates them as work progresses. This keeps critical context visible even after Auto Compact runs, letting you track progress without scrolling through the entire conversation.

Learn more →

Auto Compact (Always ON)

When context usage hits around 80%, Auto Compact automatically creates a comprehensive summary of the conversation. This preserves all decisions and code changes while freeing up space for continued work. You'll see a message when this happens. The task continues seamlessly - you don't need to do anything.

Learn more →

Context Truncation System

If your conversation approaches the model's context window limit before Auto Compact runs, Caret's Context Manager automatically truncates older parts of the conversation to prevent errors.

The system prioritizes what matters most:

• Your original task description stays
• Recent tool executions and their results remain intact
• Current code state and active errors are preserved
• The logical flow of user-assistant messages is maintained

What gets removed first:

• Redundant conversation history from earlier in the task
• Completed tool outputs that are no longer relevant
• Intermediate debugging steps
• Verbose explanations that served their purpose

This happens automatically. You'll keep working without interruption, and Caret maintains enough context to continue solving your problem effectively.

Best Practices

• Be specific - Clear objectives help Caret understand your needs
• Use @ mentions strategically - Reference specific files rather than entire folders
• Monitor the progress bar - Yellow/red means consider using /smol or /newtask
• Trust auto-management - Focus Chain and Auto Compact handle complexity automatically
• Use Memory Bank - Document persistent patterns and conventions

Next Steps

Focus Chain

Learn how Focus Chain maintains task continuity

Auto Compact

Understand automatic conversation compression

Memory Bank

Set up persistent project knowledge

Caret Rules

Define project-specific conventions