SpaceXAI introduced Grok Build memory on September 16 and Grok 4.7 on September 21, 2026. Together, the releases target longer engineering work: the model is trained for harder tasks, while the coding tool retains project knowledge across sessions.
The Grok 4.7 announcement reports stronger coding and knowledge-work results. The memory announcement describes project-scoped Markdown notes and a global preference scope. Better retained context can help an agent, but stale context can also repeatedly steer it in the wrong direction.
What Grok 4.7’s launch establishes
The release describes a larger base model and a longer reinforcement-learning run emphasizing difficult work. Its comparison table lists $2 per million input tokens and $6 per million output tokens for the standard configuration. Other access paths and modes should be checked separately.
Vendor benchmarks are useful for choosing an evaluation candidate. They do not establish the cost of completing an internal engineering task. A team should compare repository-specific results with matched instructions, tools, effort levels, and review requirements.
Memory stores knowledge that can outlive a session
Grok Build records durable conventions, decisions, and project facts in the background. Its announcement says current conversation instructions take precedence over notes, and that memory avoids secrets and information already covered by repository documentation.
The operational challenge is keeping remembered facts aligned with the code. A note about a test command may be correct today and obsolete after a tooling change. A team needs an owner and a way to inspect or remove incorrect notes rather than assuming automatic capture means automatic correctness.
Keep repository instructions canonical
Project memory is useful for continuity, but it should not become a competing policy store. Put shared requirements where the team can review and version them. Use session memory to point to that source and preserve context that is genuinely useful across tasks.
Also examine the global scope. A preference that applies to one project may be inappropriate in another, especially when security or deployment conventions differ. Concurrent agents should work in isolated task environments so one session’s unfinished edits do not become another session’s apparent baseline.
How to evaluate the combined workflow
Test repeated work across fresh sessions. Include a deliberate change to a project convention and check whether the agent follows the current source. Review what gets captured, whether sensitive material is excluded, and how a mistaken memory affects later actions.
Nerova’s assessment is that model quality and context management should be evaluated together. Grok 4.7 may improve task execution, while memory may reduce repeated setup. The gain is useful only when retained context remains inspectable and the resulting code passes the project’s actual acceptance checks.