OpenClaw keeps configuration and credentials under ~/.openclaw/ and long-term memory, persona, and skills in the workspace tree (default ~/.openclaw/workspace; multi-agent setups may use workspace-{agentId}). Community reports around 2026.3.x describe moves from a single workspace to per-agent paths where MEMORY.md and memory/YYYY-MM-DD.md were not auto-copied: Gateway still starts, but curated memory reads an empty directory—hence “the upgrade wiped the brain.” If you have not separated “config loss” from “workspace loss,” you will burn time swapping models or re-wiring channels.

The second root is scp of openclaw.json only, without credentials/, sessions/, and workspace files such as skills/, AGENTS.md, and SOUL.md; the dashboard may load while tools behave like first onboard. The third is launchd environment drift versus an interactive shell: plist WorkingDirectory or OPENCLAW_* still point at old paths after the cloud move. The fourth is skipping openclaw doctor across a large version gap, leaving schema migration half-done and fields silently reverting to defaults. Align with persistence baseline first—state directories must not live on team sync folders—or high-frequency small writes will amplify lock contention on sync layers.

After official install-daemon on the cloud host, migration acceptance should freeze four lines: workspace absolute path, openclaw version, doctor exit code, first business message id or timestamp. Any region or tier change must update one of those lines and retain the previous tarball hash for rollback—not “we probably backed up.” If channels misbehave post-migration, run the channels probe on listen ports and token truth before blaming memory directories.

Platform owners should split “migration” and “greenfield install” templates: greenfield may start from an empty workspace; migration must attach tarball checksums and doctor output. Mixing templates makes “accept empty memory” look like the default in review and leaves no story for stakeholders when curated memory vanishes.

The goal is not “tar the entire home directory” but Gateway on the target reading the same memory and skill boundaries after doctor. Must-pack items: ~/.openclaw/openclaw.json, credentials/, sessions/ when you rely on session continuity, and the full workspace tree including at least MEMORY.md, memory/, skills/, and USER.md as your team actually uses them. Optionally include seven days of rotated logs from your persistent log sink for alignment with the diagnostic ladder, but do not bloat the production migration tarball with giant logs.

Paste the table below into the change record “migration package contents” field and run it beside PATH checks from the headless SSH walkthrough: prove what is inside the tarball, then prove what the target read.

Record shasum -a 256 on the source and attach it to the change record to catch truncated SCP transfers. Move archives over approved encrypted channels (bastion rsync) rather than long-lived public object buckets holding API credentials. When workspace exceeds a few gigabytes—large skill assets or memory logs—size the target SSD tier before transfer; that ties directly to the M4 Pro decision tree below.

Replace versions with your frozen openclaw build and attach tarball shasum -a 256 to the change record. Confirm matching Node major versions on source and target; otherwise install on the target first, then unpack workspace, avoiding “new binary, old memory” half-states. If CLI is not installed on the cloud host, finish official install before unpack—do not reverse the order.

openclaw --version
ls -la ~/.openclaw/workspace*
tar -czf openclaw-migrate-$(date +%Y%m%d).tgz \
  -C "$HOME" .openclaw/openclaw.json \
  -C "$HOME" .openclaw/credentials \
  -C "$HOME" .openclaw/workspace
shasum -a 256 openclaw-migrate-*.tgz
scp openclaw-migrate-*.tgz user@cloud-mac:~/
ssh user@cloud-mac 'openclaw gateway stop || true'
ssh user@cloud-mac 'tar -xzf openclaw-migrate-*.tgz -C $HOME && openclaw doctor'

Do not gateway start immediately after extract: run openclaw doctor, save full stdout to the change record, and compare MEMORY.md timestamps with the source. When doctor suggests a workspace path migration, merge rather than overwrite unless the record explicitly accepts empty memory. Run install-daemon only after doctor passes so plist paths do not point at empty directories mid-merge.

When the six steps complete, the ticket should answer which lever moved: tarball contents, workspace path, doctor result, or tier. Even KVMNODE region-only moves should archive a tarball in artifact storage—region changes often shift IP and egress policy; rollback needs memory plus config, not a blank Gateway. During acceptance week, snapshot openclaw doctor daily and map fields to L1/L2 on the diagnostic ladder.

Workspace growth often hits disk before CPU: many memory/*.md files, skill repos, and parallel downloads stress disk IOPS and unified memory page cache. If Xcode or Flutter archive runs on the same dedicated host after migration, track root utilization after unpack, Gateway health P95, and memory pressure during parallel builds during acceptance week.

A personal Mac as the only backup source yields unpredictable backup windows with lid-close sleep and no contractible 7×24 acceptance; cron on laptops breaks on user switches and permission prompts. Sleep-heavy notebooks are poor hosts for ongoing memory log writes after migration. For teams that need auditable, region-flexible, tier-upgradable dedicated hosts for OpenClaw memory plus Gateway, KVMNODE Mac Mini cloud rental is usually the better fit: dedicated Apple Silicon, six regions, day-through-month terms, and the same placement language for “pack—unpack—doctor” and CI. Order via order page, runbooks in the Help Center, SKUs on pricing.