Teach ReacherX how you want to work

Save useful instructions, correct the audience, and understand what △ Agent can learn from later work.

SalmanFounder, ReacherX

Jul 4, 20263 min readUpdated Sep 11, 2026

You shouldn't have to make the same correction every time you use ReacherX. If you always want the first message to be short, tell it to remember that. If your workshop is free, make sure it knows that too.

You can do this by talking to △ Agent. There isn't any code to write.

Tell it what to remember

Try a request like this in the workspace where you want it used:

Remember this for outreach: keep the first message under 80 words. Ask one question. Don't ask for a meeting in the first message.

Wait for confirmation that it saved the instruction. If it only says it understands, ask it to save it as workspace memory. That makes it available for later work in this workspace.

You can give it a demo or an approved explanation too. Say what the file is for and when to use it. The attachment alone won't explain what you have in mind.

Ask for shorter introductions and check the next draft.

What if it's finding the wrong people?

That's an audience problem. Changing the writing style won't fix it.

Say you want independent designers doing client work, but it keeps finding people who sell design courses. Explain the difference and ask it to update the audience requirements. Review the proposed change before applying it.

When a requirement has changed, say which old one it replaces. Otherwise △ Agent may be trying to satisfy both. You can be direct: “We no longer need workshop speakers. We need people who want to attend.”

Does it learn by itself too?

ReacherX records events such as qualification decisions, approvals, edits, completed tasks, and replies. The memory system can use relevant events to save lessons for later work. It also notices which searches found good people and tends to run those again.

What you tell it always wins over what it guesses. And this is memory in the application, not the underlying model being retrained every time you talk to it.

You can inspect those memories and their sources in Agent observability. Read a few new matches and drafts after you change something. Are they closer to what you meant? If an instruction is outdated, ask for it to be replaced and check the saved result.

The learning system can get things wrong too. I want it to get more useful as you work with it, but you should be able to see what it remembered and where that came from. Developers can follow the memory walkthrough to see how that part is built.