Grok Build starts remembering how your team runs tests

In xAI's own example, Grok runs <em>cargo test</em>, the run fails because the PostgreSQL test database is not up, and the developer explains that <em>just test</em> should be used instead because it starts the database first. That correction now outlives the session: memory has landed in Grok Build, according to Testingcatalog, and the command survives in a testing note that Grok reads the next time it goes near tests.

At a glance

• xAI has switched on memory for Grok Build, so team conventions, technical decisions and durable project facts carry from one coding session into the next instead of dying with the terminal.
• Notes are Markdown files, one topic per subject, kept in a workspace scope per project plus a global scope for preferences, with /dream folding new observations into the topic files.
• Capture only begins after the first completed turn, and saved context applies to new sessions, so you need /new or a fresh grok before any of it takes effect.

If you missed the earlier episodes: according to ScriptByAI, Grok Build launched as a terminal coding agent for SuperGrok and X Premium Plus subscribers on May 25, 2026, was open-sourced as a coding agent harness and TUI on July 15, hit version 1.0.0 on August 7 and reached every plan across web, iOS and Android on August 19. The same outlet dates Skills, billed as reusable preferences and workflow memory across Grok apps, to May 18, 2026.

Capture runs on every completed turn, in the background

The loop is simple. After each completed turn, Grok reviews the exchange in the background and records what looks likely to stay useful: conventions, technical decisions, durable facts about the project. Nothing pauses while it does so, and the active session keeps running.

What it keeps is narrow by design: code-writing and review practices, the reasoning behind a decision, where a subsystem lives, commands such as how to run the tests. What it leaves out is just as specific: task state, tentative conclusions, secrets, and anything the repository or its docs already cover.

Think of a new colleague who does not stop the meeting to ask how things are done here, but writes the house rules in a notebook afterwards and checks them before touching that area again.

/dream files the notes, /memory shows you what stuck

Notes land as Markdown files, one topic per subject. Project work stays inside a dedicated workspace scope, while a global scope holds preferences meant to travel between projects. The /dream command folds new observations into the relevant topic files, and xAI says the same process also runs periodically on its own in the background.

/memory opens a read-only browser of those files, grouped by scope, with filtering and a preview of the selected file. From there you can copy a path, delete a note, or switch one off without removing it.

Memory is not new to Grok elsewhere in the product line. TechCrunch reported in April 2025 that xAI gave Grok chat a memory feature in beta on Grok.com and the iOS and Android apps, with xAI saying memories are transparent and users can choose what to forget, and with the EU and U.K. left out.

Notes apply before work starts, and the live conversation still wins

Before Grok begins related work, it reads the topics that cover the area it is about to touch and applies them even when the new conversation never mentions them. Instructions you give in the current conversation take priority over anything sitting in a note.

Activation carries a condition. Memory is available now, note capture begins after the first completed turn, and the saved context applies to new sessions: run /new, or start a fresh grok. New users install the tool with curl -fsSL https://x.ai/cli/install.sh | bash.

For xAI, the release turns Grok Build from a session-bound coding tool into one that develops continuity around a team and its projects, and the company says the tool should become more useful with continued use.

xAI says secrets stay out, but nothing here spells out how long notes live, where the files sit for a team rather than one machine, or what happens when two people generate contradictory conventions. Security guidance collected by Layer3Labs treats persistent agent memory as a store of prompts, code fragments and decisions that needs rules on what may enter context, how long it is retained and how it is audited. And oddly, /memory is read-only, which makes the fastest way to spot a wrong note a place where you cannot fix it.

How the topic files age

The announcement gives no retention window, no ceiling on how many notes a project accumulates, and no detail on sharing files beyond the machine that wrote them. Those are the questions a month of real use answers: whether the periodic /dream keeps a project down to a small set of readable references, or whether the topic files grow into something you end up pruning by hand.