xAI Adds Cross-Session Memory to Grok Build
The coding tool now turns durable project context into background-written Markdown notes. The practical test is whether automated recall stays accurate as work changes.
Loading page…
The coding tool now turns durable project context into background-written Markdown notes. The practical test is whether automated recall stays accurate as work changes.
Listen to this story
xAI has introduced persistent memory in Grok Build, turning selected project knowledge into inspectable Markdown files that can inform future coding sessions. The system separates project and global scopes, excludes secrets and temporary task state, and lets current-chat instructions override stored notes. Memory is available now only in new sessions, with users able to review or edit generated files through /memory.
Memory is available now in Grok Build, but users must run /new or start a fresh session to activate it.
Notes are created after completed turns and may capture conventions, decisions with rationale, stable facts, and test commands.
Task state, tentative conclusions, secrets, and information already documented in the repository are excluded.
Grok Build no longer starts every fresh coding session with a blank slate. xAI has added persistent memory that carries project conventions, decisions and durable facts into later sessions, aiming to spare developers from repeatedly explaining how a codebase works.
The feature works after a turn is complete, rather than while the user is waiting for an answer. Grok reviews the finished exchange in the background and records information xAI considers durable, without interrupting the session. That can include a team’s coding and review conventions, a decision and its reasoning, or stable facts such as where a subsystem lives and which command runs the test suite.
xAI draws a boundary around what it says the system should retain. Task state, tentative conclusions, secrets, and material already covered by the repository or its documentation are left out. The distinction matters: the feature is positioned as a record of settled working knowledge, not a full transcript or a substitute for project source material.
The retained material is kept as Markdown rather than an invisible internal store. Each project has its own workspace scope, while a global scope holds preferences that can apply across projects. Over time, the /dream command folds recent observations into one topic file per subject; xAI says that process also runs periodically in the background.
The recall step is more consequential than the file format. When a developer returns to a project, Grok reads the topics covering the area it is about to work on, including subjects that may not be raised in the new conversation. xAI says instructions in the current chat override anything stored in a note, preserving the immediate request as the controlling instruction.
Memory is available in Grok Build now, but it applies to new sessions. Users must run /new or start a fresh Grok session; notes begin after that session’s first completed turn. The release does not establish how often retained notes will be wrong or stale. It does, however, make inspection part of the workflow: users can browse the generated files and locate a note to edit when it is incorrect.
That leaves Grok Build’s next test less about whether it can save a useful convention than whether its background summaries remain dependable as a project evolves. The system’s design gives users visible files and current-session override rules, but its value will depend on whether those safeguards make accumulated context easier to trust instead of another layer developers must continually correct.
Loading discussion...
Make your case
A view to debateAutomatic memory is useful only if developers can reliably inspect and correct it.
Explain what control would make automatic memory trustworthy.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.