Grok Build Wins the Cross-Project Memory Test Against Claude Code

Grok Build had the advantage in the reported cross-project memory test: it passed all three scripted scenarios, while Claude Code passed two. Claude Code was faster overall, completing the same test setup in 66 seconds versus 165 seconds for Grok Build. The result points to a practical split rather than a universal winner: Grok Build suits workflows that need automatic global memory, while Claude Code favors speed and can use manually configured instructions across projects.

Grok Build won the cross-project test, but Claude Code was faster

The comparison used three two-session memory tests across four small Node.js repositories. Both agents passed the project-local scenarios involving test commands and coding decisions. The deciding scenario asked an agent to remember a rule when working in a different repository.

Grok Build passed 3 of 3 tests. Claude Code passed 2 of 3. The result was reported on September 21, 2026, using Grok Build 1.0.40 with Grok 4.6 at high effort and Claude Code 2.1.226 with Opus 5.

That distinction matters. Remembering a repository's conventions is one job; carrying a personal rule from one codebase into another is a broader memory workflow.

The memory systems take different approaches

xAI announced Grok Build's memory features on September 16, 2026. Its design includes separate project and global note scopes, with a /memory browser for inspecting saved information. Project notes stay tied to one codebase, while global notes can apply across projects.

Anthropic documents Claude Code's automatic memory as repository-scoped and shared across that repository's worktrees. Claude Code also supports user-level ~/.claude/CLAUDE.md and ~/.claude/rules/ files, which can apply across projects when configured manually.

That is the key distinction: Claude Code's automatic project memory is not the same thing as its user-level instruction system. A developer can use cross-project rules in Claude Code, but they are configured separately from the automatically generated repository memory.

What the three-test comparison measured

The reported setup used four small Node.js repositories and three scripted tests, with two sessions for each scenario. The agents had to recall project-local information and then handle a rule that belonged in another repository.

The tested pair was:

• Grok Build 1.0.40 with Grok 4.6 at high effort, accessed through an xAI API key.
• Claude Code 2.1.226 with Opus 5, accessed through a subscription.

The different models and access methods are important when reading the timing, token, and cost figures. They describe this particular setup, not a universal performance ranking for the products or their memory implementations.

The deciding case was a rule that had to travel

Grok Build's global memory carried a rule from one repository into a second repository during the reported test. That behavior matched xAI's documented project-and-global memory model.

Claude Code's automatically generated memory remained tied to the first repository in the same scenario. Anthropic's documented user-level ~/.claude/CLAUDE.md and ~/.claude/rules/ provide another route for instructions that should apply across projects, but that manual configuration was distinct from the automatic memory tested here.

For a developer who regularly jumps between unrelated repositories and wants one set of conventions to follow them automatically, Grok Build had the advantage in this test.

Claude Code was faster and Grok Build used fewer tokens

The reported totals show a clear speed advantage for Claude Code, alongside lower token use and lower reported test cost for Grok Build:

The $0.41 and $1.05 figures are the reported costs of those sessions under their respective model and access setups. They are not monthly subscription prices or a direct plan comparison.

Claude Code finished the three scenarios in less than half the total time recorded for Grok Build. Grok Build used fewer tokens in the same comparison, but the different models and billing paths mean the figures should be treated as measurements from that setup rather than a general cost rule.

Which agent fits your workflow?

Choose Grok Build for automatic cross-project memory

Grok Build is the more direct fit if your workflow depends on global notes that can follow you from one repository to another. Its documented memory model separates project notes from global notes, and that global scope carried the tested rule into a second repository.

Choose Claude Code for speed and explicit user-level rules

Claude Code was faster in the reported comparison. It also gives developers a documented way to apply instructions across projects through ~/.claude/CLAUDE.md and ~/.claude/rules/, although those files must be managed as user-level configuration rather than treated as automatic repository memory.

The practical choice is therefore narrow and useful: Grok Build had the edge in the tested automatic cross-project workflow, while Claude Code delivered the faster sessions and supports cross-project instructions when configured at the user level.