Garry Tan is the president and CEO of Y Combinator. He has over 738,000 followers on X. Yesterday he publicly endorsed MemPalace, calling it "impressive." In the same post, he announced GBrain, his own AI memory project.
There is one problem with the endorsement and one problem with the project.
The Endorsement
MemPalace reported Recall@5 retrieval scores as end-to-end QA accuracy. When independent developers ran actual QA evaluation, scores dropped dramatically from the reported 96.6%. The project's own GitHub issues document the discrepancies in detail (#27, #29, #39, #125, #242).
None of this is hidden. It is in the project's public issue tracker. Garry Tan either did not check, did not care or did not understand the issues.
We wrote about MemPalace's benchmarks shortly after the project first went viral: Milla Jovovich just released an AI memory system. It reached over 1.5 million people and 5,400 GitHub stars in less than 24 hours.
The Project
GBrain appeared on GitHub on April 5, 2026. Now just six days old. 43 commits. One contributor. Over 2,000 stars.
The README described three flagship features: compiled truth rewriting, a dream cycle for overnight maintenance, and entity detection on every message.
We cloned the repo and read every file. All three features are markdown documents that instruct an AI agent what to do. The codebase itself contained no rewrite logic, no scheduling, no entity detection. The words "rewrite," "stale," "synthesize," and "consolidate" do not appear in any source file. "Cron," "schedule," "setInterval," and "timer" do not appear either.
What does exist is a storage layer over PostgreSQL with pgvector, hybrid search with Reciprocal Rank Fusion, and a chunking pipeline. Reasonably competent infrastructure. But the MCP server, the primary integration point for AI agents, ships broken. The project's own issue #22 documented twelve critical bugs including race conditions, NULL embedding overwrites, and an S3 backend that a security audit note dated April 10 calls "not production-ready."
The Pattern
This is not the first time. Tan's previous project, gstack, has amassed over 69,000 GitHub stars. Developer Mo Bitar described it as "a bunch of prompts in a folder." Another founder noted that without the YC title, it would not have made Product Hunt. A developer who examined Tan's AI-generated website code found 78,400 lines including empty CSS files, duplicate assets, and test files shipped to production.
Three projects. One pattern. Big claims, big following, no independent verification.
The Stars Mean Nothing
MemPalace now has over 40,000 stars. GBrain has over 2,000 in six days. gstack has over 69,000. None of these numbers tell you whether the software works.
If you don't happen to have a Hollywood movie star friend and you aren't president of YC with 738,000+ X followers, don't worry. You can always just buy stars.
Full Investigation
We published a detailed investigation into GBrain's source code and the MemPalace endorsement on our Substack:
When the YC President Says He's "Impressed"
We are building an open benchmark for long-term AI memory, because the current ecosystem too often fails to distinguish working systems from compelling READMEs.
Top comments (0)