📄 compaction-configuration-final.md 2,771 bytes Apr 21, 2026 📋 Raw

OpenClaw Compaction Configuration — FINAL

Date: 2026-04-21 14:24 UTC
Status: ✅ Applied and Verified
Directive: Option 3 — Revert to default 20K reserve with 128K ceiling

Configuration Applied

Compaction Settings (agents.defaults.compaction)

{
  "compaction": {
    "mode": "safeguard",
    "reserveTokensFloor": 20000,
    "reserveTokens": 20000,
    "keepRecentTokens": 10000,
    "recentTurnsPreserve": 6,
    "maxHistoryShare": 0.8,
    "memoryFlush": {
      "enabled": true,
      "softThresholdTokens": 5000
    },
    "postIndexSync": "async",
    "qualityGuard": {
      "enabled": false
    }
  }
}

Per-Agent Context Tokens

Agent contextTokens
main (Wadsworth) 128000
socrates 128000
daedalus 128000

Compaction Behavior

Trigger Threshold: ~108K tokens used (128K - 20K reserve)

Reserve Allocation:
- 20K tokens reserved for Socrates's code block outputs
- Prevents hitting absolute API ceiling (128K)
- Safeguard mode preserves recent context aggressively

Why 20K:
- Socrates outputs large code blocks (Python, SQL, configs)
- 20K buffer = ~15K code + 5K tool results/response
- Prevents mid-generation truncation

Session Status

Gateway: ✅ Live (port 18789)

Current Sessions:
- Main: 127 sessions (including current)
- Socrates: 30 sessions (including quarantined)
- Daedalus: 43 sessions (including quarantined)

Fresh Sessions: New sessions will pick up configuration automatically. Existing sessions continue with previous settings until reset.

Optimizations vs Default

Setting Default Applied Effect
reserveTokensFloor 20000 20000 ✅ As specified
reserveTokens 20000 20000 ✅ As specified
keepRecentTokens ~8000 10000 +2K recent context
recentTurnsPreserve 3 6 +3 turns verbatim
maxHistoryShare 0.5 0.8 +30% history budget

Result: Same compaction threshold (~108K), but better context preservation during compaction.

Files Modified

  • /home/hoffmann_admin/.openclaw/openclaw.json
  • Added agents.defaults.compaction section
  • Added contextTokens: 128000 to all three agents

Sovereign Constraint (Active)

Per MEMORY.md directive:

"While the agents run in the cloud, the software they build must be designed for local execution."

  • Blog Content Engine → local 3080 Ti (phi4, qwen3-vl)
  • wRVU Dashboard → local models via Tailscale
  • Cloud = bootstrap; Local = target

Gateway Status

$ curl http://127.0.0.1:18789/health
{"ok":true,"status":"live"}

Configuration loaded. New sessions will use 20K reserve / 128K ceiling.

← Back