Compaction

Managing context window size through compaction.

What is Compaction?

Compaction summarizes old messages to reduce context size while preserving important information.

Why Compact?

• Token limits: Models have context window limits
• Cost: Fewer tokens = lower API costs
• Performance: Smaller contexts process faster

How It Works

Messages 1-20    →    Summary of 1-20
Messages 21-40   →    Summary of 21-40
Messages 41-50   →    Full messages

Old messages are summarized, recent messages kept in full.

Configuration

{
  compaction: {
    enabled: true,
    threshold: 50,          // Messages before compaction
    keepRecent: 10,         // Recent messages to keep full
    summaryModel: "claude-3-haiku"  // Model for summaries
  }
}

Manual Compaction

# Compact specific session
openclaw sessions compact <session-id>

# Compact all sessions
openclaw sessions compact --all

Compaction Strategies

Rolling

Summarize oldest messages as new ones arrive.

Threshold

Compact when message count exceeds threshold.

Token-Based

Compact when token count approaches limit.

See Also

• Session
• Memory