The New Stack tests Grok Build and Claude Code memory across three sessions and finds Claude's CLI forgets a rule set for all projects when moved to a second repo

Jessica Wachtel at The New Stack published a three-test memory comparison of Grok Build and Claude Code on 21 September 2026, prompted by SpaceXAI's 16 September announcement that Grok Build "keeps notes on the conventions, decisions, and project facts that come up" and reads those notes in later sessions. Claude Code has had the same feature under the name auto memory for months. Wachtel ran the CLI every user can access on a subscription.

The three tests

Each session used the tool's headless mode, which reports its own tokens and cost. Grok Build 1.0.40 ran Grok 4.6 at high effort through an xAI API key. Claude Code 2.1.226 ran Opus 5 on her subscription. Each of four small Node repos was cloned per tool. Every test used two sessions: session one planted a fact, the tool was quit, and session two gave a task where the fact mattered without mentioning it.

Test one was a test-command rule: "In this repo, npm test fails and make test passes". Test two mixed a project decision with a trap: CSV export dropped, money is integer cents, with a float helper and half-built CSV exporter left in the repo as bait. Test three stated a rule "for all my projects" in one repo, then asked for a small feature and a commit in an unrelated repo.

The results, and where the split was

Both tools passed tests one and two. Grok wrote topics/testing.md, Claude wrote orbit-api-run-tests-with-make.md, and each ran make test in session two without touching npm test. On the trap in test two, both refused the float helper and shipped a JSON export against the "dropped CSV" rule.

Test three is where they split. Grok saved the cross-project rules to a global file called git-and-code-style.md and used them in the second repo, producing the commit feat: add --help flag with usage and supported cities. Claude Code saved the same rules only inside the first repo's memory folder, warned Wachtel that its memory store "is scoped to this project's directory", and found nothing when opening the second repo. Its commit came back as Add --help flag.

The cost and speed picture

Across the three tests: Grok Build passed 3 of 3, took 165 seconds, used 390,848 tokens, and cost $0.41. Claude Code passed 2 of 3, took 66 seconds, used 576,863 tokens, and cost $1.05. Wachtel writes that the price gap "mostly reflects Opus 5 versus Grok 4.6 rather than the memory systems".

What this changes for a team choosing between them

The core promise held for both: an in-repo rule stated once was reread and followed the next session. Where the tools diverge is the scope of the note, and that scope decides whether a team can rely on the feature. A rule that only lives if a person remembers to copy it into ~/.claude/CLAUDE.md will drift out of sync the first time somebody sets up a new repo.

One test on four Node repos, with one prompt planting each rule, demonstrates behaviour under a controlled probe rather than under real workloads. Wachtel's own summary is that Grok Build "remembered everything, it carries rules across projects, and it cost less than half as much on every test", while Claude Code's Opus 5 was faster on every session. Anthropic's projects beta, announced 17 September, adds shared memory across cloud threads for select Pro and Max subscribers, so the CLI behaviour Wachtel measured may change soon.

Source

Jessica Wachtel, The New Stack, Grok Build vs. Claude Code: I tested which one has the better memory, 21 September 2026.