Grok Build vs. Claude Code: I tested which one has the better memory
The New Stack Jessica Wachtel
xAI added memory to Grok Build, and I tested it against Claude Code. Grok kept cross-project rules; Claude did not, and Grok cost less.
Based on reporting by The New Stack, Jessica Wachtel — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
xAI said on September 16 that Grok Build, its terminal coding agent, now has memory. The pitch: it can keep notes on conventions, decisions, and project facts, then read those notes in later sessions before touching related code. The notes live as Markdown files, split between a per-project workspace scope and a global scope. A /memory command browses them.
Claude Code has had something similar for months under the name auto memory. It stores a MEMORY.md index plus one file per note, per repository, and the docs say it’s on by default. Anthropic’s Projects beta, announced September 17, adds shared memory across cloud threads, but only for select Pro and Max subscribers with no existing projects. I tested the CLI versions both companies are shipping, not the demos.
I ran the same three tests on my Mac, using separate copies of four small Node repos. Each session was scripted in headless mode so I could see tokens and cost. The setup was simple: tell the tool something once, quit, come back later, and see whether it remembered without being reminded. In the first test, both tools saved the note and later used make test instead of npm test. In the second, both remembered that CSV export had been dropped and that money was stored in integer cents, not floats.
The split came in the third test. I gave each tool a rule for all projects: use conventional commit messages and don’t add comments on obvious code. Grok Build saved that rule in a global file and carried it into a different repo. Claude Code saved it only in the first repo’s memory folder, said the memory store was scoped to that directory, and then missed the rule in the second repo. It still avoided comments, but only because that’s Claude’s default.
On the numbers, Grok was the cheaper one by a wide margin. Across the three tests, it used 390,848 tokens and cost $0.41; Claude Code used 576,863 tokens and cost $1.05. Claude was faster in every session, but speed didn’t help when the memory boundary was the repo itself. For the basic claim here — remember what I said last time in the same project — both tools looked similar. For remembering across projects, Grok was the one that actually held the line.
My take — AI-written commentary, not fact-checked reporting
This is the usual AI-tools trick: call a repo boundary a feature until users trip over it. For coding agents, memory that stops at the folder is fine if you like tidy demos and useless if you work like a human across projects. Grok’s global scope is the less fancy answer, and in this case the better one.
Read more about this at: The New Stack