Y Combinator folks expected gold-standard picks from their president. Solid code, real benchmarks, the works. Tan’s tweet yesterday? Crushed that.
He called MemPalace “impressive.” Same breath, he unveiled GBrain, his stab at AI memory. Hype train leaves the station—2,000 GitHub stars in six days. But peek under the hood? Smoke and mirrors.
Look.
MemPalace went viral last month—Milla Jovovich’s glow helped, sure—racking 40,000 stars on phony Recall@5 scores masquerading as QA accuracy. Indie devs tested it properly. Scores tanked from 96.6%. GitHub issues scream the truth (#27, #29, #39, #125, #242). All public. Tan either skipped homework, shrugged, or missed the plot.
What Did Garry Tan Actually Tweet?
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.
That’s the original post’s hook. Impressive, huh? Blind endorsement from a YC kingpin. Stars explode. Repos balloon. But code? Nah.
We cloned GBrain. April 5 drop, 43 commits, solo author. README boasts compiled truth rewriting, dream cycles for overnight tweaks, entity detection per message. Sounds dreamy.
Reality check: Those “features”? Markdown docs bossing around some AI agent. No rewrite code. Zip. Words like “rewrite,” “stale,” “synthesize,” “consolidate”? Absent from source. No cron jobs, no setInterval, no timers. It’s prompts in a folder—déjà vu.
What is there? Decent Postgres/pgvector storage, RRF hybrid search, chunking pipe. Competent infra, if you’re generous. But the MCP server—the AI agent hookup? Doomed. Issue #22 lists twelve bugs: race conditions, NULL embedding nukes, S3 backend flagged “not production-ready” in an April 10 audit note.
Broken on arrival. Ship it anyway?
This ain’t Tan’s debut rodeo. Gstack: 69,000 stars. Mo Bitar called it “a bunch of prompts in a folder.” Sans YC badge, it’d flop on Product Hunt. His AI-site generator? 78,400 lines of bloat—empty CSS, dupes, prod-shipped tests.
Pattern screams loud. Big claims. Massive followings. Zero verification.
Why Did Garry Tan Endorse MemPalace’s Bogus Benchmarks?
Maybe rush to launch GBrain. use the buzz. Or genuine blind spot—YC echo chamber where stars > substance. Here’s my hot take, absent from the original scoop: This reeks of 1990s vaporware, pre-dotcom crash. Remember Pets.com? Flashy sites, zero backend. Tan’s repos are that—README unicorns pulling dev carts blindfolded. Prediction: Open-source AI memory field fractures. Trust erodes. Devs flock to verified benches like the one they’re building on Substack.
And buying stars? Cheapest hack in town. No Hollywood pal or 738k followers? Wallet to the rescue.
MemPalace hit 1.5 million eyeballs off Jovovich. GBrain rides Tan’s coattails. Gstack coasts on legacy. Stars don’t debug race conditions. Or fake metrics.
But here’s the kicker—it’s all open. Fork it. Fix it. Or walk away.
Tan didn’t invent hype. YC didn’t corner clout-chasing. Still, when the president’s batting average is prompts-over-product, it stings. Devs deserve better than celebrity repos.
Skeptical? Clone ‘em yourself. We did. Truth’s in the commits.
Is GBrain Worth Your Time as a Developer?
Short answer: For infra plumbing, maybe. Fork the storage layer—it’s solid. Rest? Vapor. No auto-maintenance, no real entity magic. Agents gonna choke on that MCP mess.
Why care? AI memory’s hot. Long-term recall could flip agents from forgetful pets to workhorses. But hype drowns signal. Tan’s spin—“flagship features” via docs—sets bad precedent. Corps copy it. Startups ape. Result? GitHub graveyard of half-baked stars.
Unique angle: Parallels Theranos blood tests. Big names (Tan = Holmes?), wild promises, no independent audits. AI winter 2.0 incoming if unchecked.
Fix? Demand benches. Real QA evals. Not README fairy tales.
🧬 Related Insights
- Read more: Stack Overflow’s 25 Million Questions: The AI Data Gold Rush You Can Mine Today
- Read more: Power BI’s Data Frankenstein: Stitching Sources Together Without Screaming
Frequently Asked Questions
What is GBrain AI memory system?
GBrain’s a GitHub repo by YC’s Garry Tan claiming AI agent memory with rewriting, dream cycles, entity detection. Reality: prompts telling agents what to do, plus basic vector DB—no actual code for core claims.
Did Garry Tan know MemPalace benchmarks were fake?
Public GitHub issues detailed the Recall@5 vs. real QA drop months ago. Tan endorsed anyway—didn’t check, didn’t care, or didn’t get it.
Is GBrain production ready?
Nope. MCP server has 12 bugs per its own issue #22, including races and unready S3. Fork at your peril.
Will YC projects always be hype over code?
Pattern holds for Tan’s trio. Broader YC? Jury’s out—but verify before starring.
