Source - Compaction (OpenClaw Docs)

Origin: https://docs.openclaw.ai/concepts/compaction
Fetched: 2026-04-15

Summary

OpenClaw's compaction system automatically summarizes older conversation turns when sessions approach a model's context window limit. Compaction is a durable, transcript-level operation — it permanently saves condensed summaries to the session file while preserving the full original history on disk, distinguishing it from per-request session pruning which silently drops tool outputs.

Key Points

• Trigger conditions: context limit approach, model overflow errors (request_too_large, context length exceeded, etc.)
• Tool-pair preservation: tool calls remain paired with results during summarization
• Memory flush: agents can flush critical info to persistent storage before compaction runs
• Config path: agents.defaults.compaction in openclaw.json
• Configurable model: agents.defaults.compaction.model uses provider/model-id format — can point to a separate (more capable) model or local Ollama
• Custom providers: plugins register via registerCompactionProvider(), referenced with provider: "my-provider"
• Identifier policy: identifierPolicy: "strict" (default) | "off" | "custom"
• User notification: notifyUser: true displays status messages at compaction start
• Status indicators: verbose mode → 🧹 Auto-compaction complete; /status → 🧹 Compactions: <count>
• Manual trigger: /compact [guidance] with optional emphasis instructions
• Escape hatch: /new starts a fresh session entirely

Linked Wiki Pages

• Compaction — full concept page