Context Compaction

Carry Code’s intelligent context compression feature automatically manages conversation length, keeping you efficient during long conversations.

---

Why Do We Need Compression?

AI conversations have an important limitation: Token count. Each AI model has a maximum token limit (e.g., 128K, 200K, etc.).

When conversations get long:

• Exceeding the limit prevents continued conversation
• Sending long contexts increases API costs
• Affects AI response speed

Carry Code’s intelligent compression solves these problems.

---

How It Works

Automatic Compression

Carry Code automatically detects conversation length and when approaching the token limit:

1. Analyze conversation - Identify key and secondary information
2. Compress secondary content - Simplify early detailed conversations
3. Preserve key information - Keep important code and decisions
4. Maintain coherence - So AI still understands the context

Compression Strategy

Content Type	Handling
Key code	Fully preserved
Important decisions	Fully preserved
Error messages	Condensed
Casual chat	Heavily simplified
Historical details	Selectively compressed

---

Manual Trigger Compression

Command

/compact

Manually trigger context compression.

Use Cases

• When conversation gets too long
• Want to clear unnecessary history
• Want to save API costs

Compression Effect

After compression:

• Token count significantly reduced
• Key information still preserved
• Conversation can continue

---

Configuring Compression Behavior

Configuration File

Configure in ~/.carry/carrycode.json:

{
  "compaction": {
    "enabled": true,
    "threshold": 80000,
    "preserveKeyInfo": true
  }
}

Parameter Description

Parameter	Description	Default
enabled	Enable automatic compression	true
threshold	Token threshold to trigger compression	80000
preserveKeyInfo	Preserve key information	true

---

Impact on Conversations

What Happens

1. Information simplification - Early detailed conversations are condensed
2. Key points preserved - Important code and decisions are kept
3. Continue chatting - You can continue the previous task

What Doesn’t Happen

• ❌ Won’t lose key information for current task
• ❌ Won’t delete important code
• ❌ Won’t affect current conversation context

---

Usage Tips

1. Regular Compression

When conversation exceeds a certain length, proactively use /compact:

• Keep conversations efficient
• Save costs
• Avoid hitting token limits

2. Create New Sessions When Appropriate

For completely different tasks, creating a new session is better:

• Avoid context confusion
• Keep each task clear

3. Use Plan Mode

When you don’t need to write code, use Plan mode:

• Won’t generate new code content
• Context is simpler

---

Combined with Session Management

Operation	Effect
/session new	Brand new blank context
/compact	Compress current context
/session switch	Switch to other context

---

FAQ

Q: Will compression lose important information?

No. The compression algorithm prioritizes preserving:

• Key code snippets
• Important decisions
• Current task-related information

Q: Can I continue previous work after compression?

Yes. AI will understand the compressed context, and you can continue previous tasks.

Q: How to disable automatic compression?

Set "enabled": false in the config file, but it’s recommended to keep it enabled to avoid hitting token limits.

---

Next Steps

• Dual Mode - Use with modes
• Session Management - Manage multiple sessions
• Approval Modes - Control AI permissions