Decisions
This system is the result of deliberate choices. Local work stayed local. Memory moved remote. Private services stayed private. GitHub stayed separate from the memory layer. The point of this page is to make those choices explicit and explain what each one solved.
1. Keep active work local
Editing files, exploring repos, and running coding agents are all better locally. The feedback loop is faster, local tools are easier to control, and interactive work stays close to the real code.
This decision solves the speed problem. It keeps day-to-day work responsive and avoids turning every code change into a remote systems problem.
2. Move persistence remote
Memory and orchestration need to outlive local sessions. The VPS is the right place for that because it stays available even when the laptop is closed, rebooted, or disconnected.
This solves the continuity problem. Without it, the system always risks collapsing back into local-session amnesia.
3. Use both Claude and Codex
The environment keeps both tools because they are useful in different ways. Claude is strong for guidance, explanation, cleanup, and structure. Codex is strong for implementation-heavy repo work.
This solves the tool-fit problem. The setup does not force every task through one agent when the work benefits from more than one.
4. Split memory from source control
Git stores code, workflows, docs, and publishable files. OpenClaw memory stores decisions, rationale, and cross-session context. That distinction is structural.
This solves the overload problem. If everything goes into git, the reasoning disappears. If everything goes into memory, the code loses a clean version boundary.
5. Use hooks for ingestion
Commits and session endings feed the memory layer through hooks and helper scripts because manual logging alone is fragile.
This solves the capture problem. Useful context gets preserved even when the user or the agent does not stop and write notes manually.
6. Keep internal services private
The memory layer is not a public app. Tailscale, loopback binding, SSH tunneling, and the firewall are part of the design, not cleanup work that gets added later.
This solves the exposure problem. Internal systems stay private and the public edge stays small.
7. Keep the public site static
The site is documentation. It does not need a custom runtime stack on the VPS. Keeping it static lowers maintenance cost and keeps the public-facing layer simple.
This solves the complexity problem. The VPS can focus on OpenClaw and the private orchestration layer instead of also hosting the docs site.
8. Use plain Markdown as the memory substrate
The workspace depends on Markdown files because they are easy to read, version, diff, and reason about. There is no need for a fancy hidden format when the goal is continuity and inspection.
This solves the transparency problem. The memory layer stays visible and understandable.
Decision pattern