2026 OpenClaw Active Memory & Sub-Agents on Remote Physical Macs: MEMORY.md Migration, Write-Heat & JSONL Audits Aligned to Gateway Sandbox—7×24 Long-Session Runbook (Config Snippets + FAQ)

2. Pain points

1. Constraint: relative paths × multiple cwds. Sub-agents started from a repo subfolder write MEMORY.md deep in that subtree while the parent still reads the old location—path split, not mysterious amnesia.
2. Hidden cost: write-heat is write amplification. Fsync-heavy full rewrites each turn, combined with audit JSONL, can peg iowait on 7×24 hosts before business traffic does.
3. Stability & audit: two writers, two trees. When the sandbox blocks host paths, sub-agents land audit files under private dirs and host jq joins fail—usually requestId and writable-path policy were never unified.

3. Decision matrices

Use these before rollout to agree where memory lives and how aggressively you flush to disk.

4. Config snippets (sandbox-aligned)

Illustrative JSON—field names must match your pinned OpenClaw version. The non-negotiables are absolute paths, allowlists, and requestId propagation.

{
  "memory": {
    "path": "/Users/ocagent/workspace/MEMORY.md",
    "encoding": "utf-8",
    "flush": { "maxEvents": 32, "intervalMs": 5000, "halfLifeRounds": 12 }
  },
  "subagents": {
    "defaultCwd": "/Users/ocagent/workspace",
    "inheritEnv": ["OPENCLAW_REQUEST_ID", "MEMORY_PATH"]
  },
  "gateway": {
    "auditJsonl": "/Users/ocagent/logs/gateway-audit.jsonl",
    "sandbox": {
      "writable": [
        "/Users/ocagent/workspace/.sandbox/scratch",
        "/Users/ocagent/logs/sandbox-audit.jsonl"
      ]
    }
  }
}

In your launchd plist, export MEMORY_PATH and OPENCLAW_REQUEST_ID (or have a gateway wrapper set them per request) so SSH foreground runs do not diverge from unattended daemons.

5. Seven-step runbook (remote physical Mac)

1. Freeze authoritative paths. Document MEMORY file, audit JSONL, and sandbox scratch as absolutes on the runbook cover—no split between wiki and code comments.
2. Migrate MEMORY.md. Copy under maintenance or read-only mode, verify with diff, stamp migratedAt metadata in the header block.
3. Normalize sub-agent cwd. Force defaultCwd to the main workspace or pass an explicit memory flag equivalent.
4. Turn on JSONL schema. Agree on ts, level, requestId, actor, memoryOp, durationMs for gateway and children.
5. Sandbox acceptance. Probes: writes outside the allowlist fail and emit deny lines; successes stay inside writable.
6. Tune write-heat. When moving from per-turn full flush to windowed batches, replay a 24 h conversation script and watch iostat plus JSONL growth.
7. Replay join. Outer-join host and sandbox JSONL on requestId with jq; any missing pair blocks production promotion.

6. Cite-ready thresholds

• Batch flush starter window: about 3–8 s or 16–32 events, whichever hits first; lengthen with half-life for very long sessions.
• Audit JSONL rotation: split near 128–256 MB so object-storage sync does not copy multi-gigabyte tails.
• Join SLA: in pre-prod soak, 0% missing requestId pairs for one hour; production alert if sustained above 0.1% excluding clock-skew buckets you document.

7. FAQ

8. Summary: why Mac mini fits long-session memory + audit IO

Active Memory with JSONL auditing is effectively high-frequency small IO plus long-lived processes: disk quality, scheduler fairness, and power behavior determine whether your write-heat strategy actually sticks. Apple Silicon Mac mini (for example M4) pairs unified memory bandwidth with roughly ~4 W class idle power, which is ideal for silent 7×24 gateways that keep memory and logs on fast local storage.

macOS also gives you launchd, Unified Logging, and APFS snapshots so triage paths stay short when a sandbox deny line or a burst flush regression appears—complementing Gatekeeper and SIP hardening versus typical Windows malware exposure on always-on boxes.

If you are moving long sessions from a laptop to a remote physical Mac so sub-agents, gateway sandbox, and JSONL rotation share one predictable baseline, Mac mini M4 is one of the most cost-effective anchors today—then fix paths once and let the same hardware profile carry your runbook through staging and production.