xAI adds persistent memory to Grok Build for cross-session project context

xAI announced on September 16, 2026 that Grok Build, its coding agent, now has persistent memory: it automatically records project conventions, decisions, and facts as it works, then reads those notes back before touching related code in a future session.

What's new

• "Grok Build now has memory. As you work, it keeps notes on the conventions, decisions, and project facts that come up, and later sessions read those notes before touching related code," xAI wrote in the announcement.
• Note-taking happens in the background as Grok completes turns, without interrupting the current workflow.
• Notes are stored as markdown files, one topic per subject — for example, code style, testing procedures, or architectural decisions — rather than a single unstructured log.
• A new /memory command lets a user browse all stored notes, organized by scope, with a searchable preview.
• A new /dream command automatically organizes recent notes into topic files, consolidating what's accumulated during a work session.
• xAI is explicit about precedence: "Instructions in the current conversation take precedence over anything in a note," so a stored note cannot override what a user actually tells Grok in the moment.
• Memory is live now and applies to new sessions going forward.

Context

Coding agents that run long sessions against the same codebase face a recurring problem: without persistent memory, every new session starts from zero, forcing a developer to re-explain conventions, past decisions, and project-specific quirks each time. Competing agentic coding tools have been converging on some form of durable, session-spanning memory to address exactly this, typically as markdown or note-style files that get consulted before new work starts. Grok Build's implementation follows that same shape — background note-taking plus an explicit review-and-organize step — rather than introducing a fundamentally different memory architecture.

Why it matters

For teams already using Grok Build on a real codebase, memory should cut down on repeated re-explanation of the same conventions and past decisions across sessions, which is one of the more common complaints about agentic coding tools once a project gets past its first few days. The explicit precedence rule — current instructions beat stored notes — is a deliberate safety valve: it means a stale or wrong note can't silently override what a developer is telling the agent right now, which matters once notes accumulate over weeks of work. Whether this meaningfully changes how teams use Grok Build depends less on the feature existing than on how accurately the automatic note-taking captures the things that actually matter on a given project, which will only become clear with real usage.