Your assistant is already working when another thought occurs to you. You ask it to recall a decision, but it cannot find it. Or a tool is installed and still does nothing. These small moments determine whether an AI setup feels useful day to day.

We have run into all of these while building our own OpenClaw setup. Three things helped: adding instructions during a task, making memory search understandable, and testing computer use beyond its installation screen. Here is what matters, including a few detours you can avoid.

1. Say it while your agent is still working

Your agent is planning the week. You remember that Friday needs to stay free. Add that instruction to the same conversation in Miracles.

OpenClaw calls this steering: a new message becomes part of the work already underway. It is the default in the documentation checked for this article. A session configured differently, or a different runtime, may instead hold messages until the next response.

In our setup, steering was already active. The useful step was a practical test, not another setting: start a plan, add “Keep Friday free” while it is being prepared, and check the result.

If your addition arrives later, ask your agent to inspect that specific session's mode. This belongs to OpenClaw and its runtime, not a switch in Miracles. A tool already running may still finish. For an important correction, ask the agent to confirm what it will now do differently.

2. A failed search does not mean your notes are gone

“What did we decide about typography?” should find a note saying “no external fonts”. Semantic search helps with this: it finds related meanings, not just matching words. The numerical representations it uses are called embeddings.

First, there needs to be a saved decision. Try asking: “Record that this project should use no external fonts, and show me where you saved it.” Later, ask about it in different words and request the source.

Our detour: treating a timeout like lost memory

One of our search failures came from a time limit in the tool connection. The index and saved notes were intact when checked. Searching through OpenClaw's existing native CLI — its command-line interface — worked. Rebuilding everything would have missed the actual problem.

Our advice: First distinguish an empty result from a search that never finished. Have the agent inspect the exact error, provider and index status. A timeout and a search with no matches are different findings. Do not delete memories or switch providers on a hunch.

Another issue appeared after we expanded the search sources: OpenClaw reported index sources changed. Its native index reconciliation resolved that. Separately, an ongoing session still saw the old tool state while a fresh check could already search the new sources. When settings and behaviour disagree, comparing with a fresh session can help.

For your setup, start by checking which files and conversations are included. An external embedding provider receives the text it processes; local embeddings perform that step on your hardware. The location of the index alone does not tell you where processing happens.

3. Installed is not the same as working computer use

Computer use lets an agent inspect and operate an interface: open a website preview, click a button and check the result. Files and connected apps often have more direct tools. Screen interaction is especially useful when the visible interface is itself part of the task.

After a change to our desktop integration, the computer-use plugin was still present but its connection to the tool service was missing. It looked installed and was nevertheless unavailable. In a separate case, the way the background process was launched mattered. Repeatedly toggling the same macOS permissions would not have been a reliable diagnosis.

Check these three layers separately:

  1. Installation: Is the intended integration present and supported by your version?
  2. Connection: Can the current session actually reach its tool service?
  3. Interaction: Can it see the intended app and perform a harmless action?

A local test page with an expandable section makes a useful first check. Ask the agent to describe the page, open the section and inspect it again. That tells you more than a green plugin list.

OpenAI computer use and OpenClaw's own device connection are different paths. Success or failure on one does not prove anything about the other. When troubleshooting, ask which path is being tested. If an approval does not arrive, check the approval flow in the client you are using before changing system permissions.

A focused check for your own setup

This task lets your agent inspect those three areas in your actual installation. You do not need to copy configuration keys from an article.

Inspect my existing OpenClaw setup for use with Miracles, read-only first. Identify the installed version and runtime, and use their current documentation.

Check three things: How do additions reach an ongoing task? Which sources does memory search cover, and are the index and embedding provider reachable? Is computer use merely installed, or actually connected in this session?

If something fails, distinguish settings, connection and functional failures. Preserve the exact error without credentials. Compare stale tool state with a fresh check before proposing a repair. Do not delete memories, install anything, or change permissions or public access yet.

Briefly explain what already works and which single next step would help most. Give its scope, a way back and a harmless functional test. Also check which text or screenshots go to external providers and what costs matter. Wait for my selection before making changes.

Start with what actually gets in your way

If additions already arrive during a task, steering is done. If search returns the right decision with its source, you do not need another search provider. If most of your work involves text and files, computer use can wait.

For a first project, try these three ideas. If connecting from your iPhone is the problem, begin with Miracles setup: test on the computer, then on your phone, then over mobile data.

Further reading

The experiences above come from our own setup; those historical issues may already be fixed in your version. For specific settings, use current documentation, last checked for this article on 13 November 2026: