Run /new inside Grok Build — or start a fresh grok instance — and memory switches on for that session automatically: the grok build memory feature reached general availability in version 1.0.34 on 16 September 2026, per the official SpaceXAI announcement, and from the first completed turn it quietly records the conventions, decisions and project facts that come up in your work. Later sessions read those notes before touching related code, which means the tool stops meeting you as a stranger every morning. This page covers exactly what it records, where the notes live, the two commands that control it, and how it compares to the file-based memory setups agents like Hermes have run for months.
📺 Watch: Claude Desktop + Obsidian + Agent OS AI Memory Systems!
🔥 Get the Agent OS as a free bonus: AI Profit Boardroom members get the full Agent OS zip, prompt libraries, daily tutorials and weekly live coaching calls. → Get inside · Want AI SEO help 1-on-1? Book a free SEO strategy session →
Every claim here is drawn from two sources published by SpaceXAI: the memory announcement on x.ai dated 16 September 2026, and the Grok Build changelog entry for version 1.0.34 of the same date, which marks memory as generally available alongside a fix for Markdown heading theme colours.
What The Grok Build Memory Feature Records
The design goal, per the announcement, is durable information rather than a transcript. After every completed turn, Grok reviews the exchange in the background and writes down what is likely to remain useful: the conventions your project follows, the decisions you have made, and the facts about the project that keep coming up. Three categories are explicitly excluded — task state, tentative conclusions, and secrets. That exclusion list is worth pausing on, because it addresses the classic failure modes of agent memory: half-finished work getting frozen as fact, speculation hardening into policy, and credentials leaking into notes.
The announcement is also specific about when capture happens: it runs after a turn completes and never blocks the session. You do not wait while the model takes notes; the recording happens behind you. Notes begin after the first completed turn of a fresh session, so the feature applies to new sessions rather than retroactively rewriting old ones.
Scopes: Workspace Notes And Global Preferences
Memory in Grok Build operates at two levels. The workspace scope holds project-specific knowledge — the conventions and decisions tied to the code in front of you. The global scope holds your own preferences, the things that should follow you across every project. The split matters in practice: a naming convention belongs to one repository, but your preference for terse commit messages belongs to you, and mixing the two is how memory systems become noise.
Under the hood, notes are stored as markdown files organised by topic. Before starting related work, Grok reads the applicable topics and applies them — with one clearly stated precedence rule from the announcement: instructions in the current conversation take precedence over anything in a note. So a stale note cannot overrule what you just said, which is the correct hierarchy and not one every memory implementation gets right.
If you want agents that compound context into actual output — content, SEO and automation systems that get sharper every week — the AI Profit Boardroom gives you the Agent OS bonus, 1,000+ workflows and weekly live coaching to build them → check out the AI Profit Boardroom. Want a 1-on-1 roadmap for your setup? Book a free SEO strategy session.
The Two Commands: /memory And /dream
Two slash commands control the system, and both are described in the announcement:
| Command | What it does |
|---|---|
| /memory | Opens a read-only browser of every memory file, grouped by scope, with a preview of the selected file — the fastest way to see what a session produced |
| /dream | Organises recent notes into topic files on demand; the same organisation also runs periodically in the background |
The /memory browser being read-only is a deliberate choice worth noting. You inspect what the system believes through the browser, and the note files themselves are markdown organised by topic. The /dream command — the announcement's name for consolidation, not ours — is the tidying pass: it takes the raw recent notes and files them into coherent topics, either when you invoke it or on its own schedule in the background. If you have ever watched an agent's notes pile up into an unusable heap, you will recognise exactly which problem this is solving.
How To Start Using The Grok Build Memory Feature Today
Getting onto the feature takes three steps, all from the announcement and changelog. First, be on Grok Build 1.0.34 or later — that is the version where memory went from preview to generally available on 16 September 2026. Second, start a fresh session: run /new or launch a new grok instance, because memory applies to new sessions. Third, just work — notes begin after the first completed turn, with no setup ritual beyond that. After a session or two, open /memory and read what it captured; that inspection habit is how you catch a wrong note early, and the precedence rule means you can always override a note by simply saying so in the conversation.
What the announcement does not specify is also worth knowing: it does not state pricing or plan availability for the memory feature, and it does not document a disable switch or privacy controls in the launch post. If either of those matters to your setup, check the current Grok Build documentation before rolling it out across a team, and treat the exclusion of secrets from capture as a design intention to verify rather than a guarantee to lean on.
How It Compares To File-Based Agent Memory
If this all sounds familiar, it should. Markdown files organised by topic, read before work starts, written after work completes — that is the same architecture the Hermes memory system has used all along, with soul.md and daily notes on your own disk, and it is the pattern the OpenClaw memory persistence setup follows too. The convergence is the story: the major agent tools are all landing on plain-text, topic-organised memory because it is inspectable and portable in a way opaque memory databases never were.
The differences sit at the edges. Grok Build's version is automatic and curated for you — capture after every turn, background consolidation via /dream, a read-only browser. The Hermes approach hands you the files directly to edit at will. Which trade-off you prefer depends on whether you want memory as a managed service inside your coding tool or as an asset you own and tend yourself. For agent skills that lean on accumulated context either way, the best Hermes agent skills round-up shows what compounding memory actually unlocks, and the Goldie Bench write-up tracks how the big agent brains compare in hands-on tests if you are deciding which model deserves your accumulated context in the first place.
Zoom out and memory is one pillar of a larger pattern — the Agent OS approach treats shared memory as load-bearing infrastructure alongside orchestration and mission control, whichever vendor's implementation you standardise on.
📺 Watch: NEW Grok Update is AMAZING!
A Busy Week For Agent Tooling
The memory GA landed in the middle of a dense stretch of agent-tool news: OpenHands shipped agent-profile secrets scoping in its 1.20.0 release the following day — covered in the OpenHands 1.20 release notes — and it arrived barely a week after DeepSeek's multimodal small-model push, covered in the DeepSeek V4.1 Flash vision capabilities write-up. The through-line across all three: the competition in agent tooling has moved from raw model quality to operational features — memory, permissions, persistence — which is precisely the layer where working professionals feel the difference daily. If you are building your stack around one coding agent, these operational releases are the ones to track — the learn Claude Code guide makes the same point from the Anthropic side of the fence.
Grok Build Memory Feature FAQs
What is the grok build memory feature?
It is Grok Build's built-in note-taking system, generally available since version 1.0.34 on 16 September 2026. After each completed turn it records durable project conventions, decisions and facts as markdown files organised by topic, and later sessions read the applicable topics before starting related work.
How do I turn on memory in Grok Build?
Update to version 1.0.34 or later, then start a fresh session with /new or a new grok instance. Memory applies to new sessions automatically, and notes begin after the first completed turn — there is no separate enable step described in the announcement.
Does Grok Build memory store secrets or passwords?
The announcement states that secrets are excluded from capture, along with task state and tentative conclusions. It does not document additional privacy controls in the launch post, so verify against the current documentation before relying on that behaviour in sensitive environments.
Can memory override what I tell Grok in a conversation?
No. The announcement states that instructions in the current conversation take precedence over anything stored in a note, so a stale memory cannot overrule a live instruction.
What does the /dream command do?
It organises recent notes into topic files on demand. The same consolidation also runs periodically in the background, so /dream is there for when you want the tidying pass to happen now rather than later.
Verdict: Memory Went From Party Trick To Default
General availability is the real headline here. Session memory in coding tools has spent a year as a preview feature you enabled at your own risk; as of 16 September 2026, Grok Build treats remembering your project as the default behaviour of a fresh session. The implementation makes sound choices — non-blocking capture, secrets excluded, conversation-over-notes precedence, markdown you can inspect — and the open questions it leaves are documentation gaps rather than design flaws. If your daily tool is Grok Build, start a fresh session and let it begin compounding; if it is not, expect your tool to copy this within a quarter, because memory has just become table stakes.
If you want the full compounding stack — agents with memory, skills and systems that get more valuable every week you run them — the AI Profit Boardroom includes the complete Agent OS zip, prompt libraries and five live coaching calls a week → get inside the AI Profit Boardroom. Rather start with a 1-on-1 conversation? Book a free SEO strategy session.
