Summary

Currently agents.defaults.compaction applies uniformly to all agents. In multi-agent deployments, agents have fundamentally different context-retention needs. An orchestrator agent needs long memory; a specialist agent should compact aggressively after task completion.

The Problem

We run 17 agents in a 4-tier hierarchy:

But OpenClaw only supports:

{
  "agents": {
    "defaults": {
      "compaction": {
        "mode": "default",      // Same for ALL agents
        "reserveTokens": 12000  // Same for ALL agents
      }
    }
  }
}

Result

• Orchestrator compacts too aggressively → loses dispatch history → forgets what it assigned to whom
• Specialist agents compact too slowly → waste context window on completed tasks → eventually hit token overflow

Proposed Solution

Per-agent compaction override

{
  "agents": {
    "defaults": {
      "compaction": {
        "mode": "default",
        "reserveTokens": 12000
      }
    },
    "entries": {
      "main": {
        "compaction": {
          "mode": "safeguard",           // More conservative for King
          "reserveTokens": 20000,
          "thresholdPercent": 70          // Only compact at 70% full
        }
      },
      "dev-lead": {
        "compaction": {
          "thresholdPercent": 65
        }
      },
      "dev-frontend": {
        "compaction": {
          "mode": "default",
          "reserveTokens": 8000,
          "thresholdPercent": 55          // Compact early and often
        }
      }
    }
  }
}

Why This Matters

1. Different agents have different memory needs — This is fundamental to any multi-agent system
2. No workaround exists — We tried building external compaction via session-guardian skill, but compaction is a core engine feature; external tools can only trigger it, not configure it
3. Prevents the fix: add @lid format support and allowFrom wildcard handling #1 cause of "agent amnesia" — Wrong compaction settings = lost context = repeated work = frustrated users
4. Small implementation surface — The compaction config already exists at agents.defaults; extending it to agents.entries.<name>.compaction follows the existing override pattern

Our Workaround (Published on ClawHub)

We built session-guardian with a "smart compaction" layer:

1. First-time install: LLM analyzes each agent's role → generates compaction-strategies.json
2. Cron runs auto-compact.sh hourly → applies per-agent thresholds externally
3. Zero ongoing token cost (config generated once, executed by scripts)

This works but is inherently limited — external tools can't control the actual compaction algorithm, only monitor and react.

Validation

We tested all 17 agents with personalized strategies. Results:

• Token usage reduced ~30-50% across the fleet
• Orchestrator (King) retained dispatch history through compaction
• Specialists compacted faster, improving response speed
• But: still can't control the actual compaction behavior, only the triggering threshold

Related

• Feature: Native skill restore after context compaction #28367 — Native skill restore after context compaction
• Safeguard compaction fails to fire when Anthropic prompt caching masks true context size #30135 — Safeguard compaction fails when prompt caching masks true context size
• [Bug]: direct session can exceed configured contextWindow without timely compaction #50065 — Session can exceed contextWindow without timely compaction