On September 16, 2026, xAI announced a long-term memory capability for Grok Build, its terminal-based coding agent, allowing the tool to carry knowledge forward between separate sessions. The system automatically records project conventions and design decisions in the background and draws on them in later work, offering one answer to a complaint developers have voiced across the AI coding tool industry: having to re-explain the same context every time a conversation restarts.
Notes Written After Every Conversation
Grok Build's long-term memory kicks in once a conversation wraps up. A background process reviews the exchange and pulls out details likely to matter later, such as coding conventions, design decisions and the reasoning behind them, and how to run the test suite. These are saved as Markdown notes, while task-specific state, tentative conclusions, secrets, and anything already documented in the repository are left out.
Notes are organized into two scopes: one specific to each project, and a global scope shared across all of a user's work. The project scope is meant to hold conventions and specifications unique to that codebase, while the global scope covers preferences that apply no matter which project is open.
When a new conversation starts, Grok reads the notes relevant to the task at hand before getting to work. It isn't recalling the original conversations verbatim, only the distilled notes drawn from them. If a note ever conflicts with what the current conversation asks for, the live instructions take priority.
Two Commands for Managing the Notes
Developers can inspect and organize what has accumulated. The /memory command opens a read-only browser that lists notes by project and global scope, with previews of each file. The /dream command consolidates scattered observations into a single topic file; it can be run manually or left to fire automatically in the background on its own schedule. A separate manual command is also available for jotting down a note on the spot.
Grok Build currently runs on xAI's Grok 4.6 model, and the memory feature is already active for every user's new sessions. No opt-in or configuration change is required — recording simply begins the next time a conversation starts.
Developers who rely on AI coding agents day to day frequently point to the same frustration: having to restate a project's context, and any corrections already given, every time a session resets. Grok Build's long-term memory is aimed squarely at that friction, but xAI isn't alone in tackling it.
In March 2026, GitHub turned on Copilot Memory by default for Pro and Pro+ subscribers, a feature that automatically detects and stores a repository's conventions and design patterns. Unlike Grok Build, however, it is built to expire stored facts and preferences automatically once they go unused for 28 days, to keep the information accurate. Anthropic's Claude Code takes yet another approach, combining developer-written instruction files with an automatic note-taking system that captures corrections and preferences as they come up. Each company is drawing its own line on what gets remembered and how much of that process should run automatically.
Summary
xAI has added a long-term memory feature to its Grok Build coding agent that automatically extracts key details from conversations and stores them as Markdown notes, rolling it out to every user's new sessions as of September 16, 2026. The goal is to cut down on re-explaining project context session after session, and rivals such as GitHub Copilot and Claude Code are tackling the same problem with their own designs, underscoring how memory has become a defining factor in how usable an AI coding agent feels.
