Former Coinbase Engineers Built an AI That Remembers Everything You've Ever Decided
The most unsettling thing about Rowboat isn't that it can build presentations from your scattered work conversations—it's that it probably knows your actual priorities better than you do.
This Y Combinator S24 startup, launched February 10th by former Coinbase engineers who worked on graph neural networks, has created something that sounds mundane but feels revolutionary: an AI that actually remembers your work.
<> "AI agents that can run tools on your machine are powerful for knowledge work, but they're only as useful as the context they have."/>
That context problem is everywhere. Your AI assistant doesn't know that you promised Sarah a deck by Friday. It has no idea that the budget discussion from three weeks ago is now blocking the roadmap you're trying to build today. Every interaction starts from zero.
Rowboat fixes this by doing something creepy-smart: it reads your Gmail and meeting notes from services like Granola and Fireflies, then builds a persistent knowledge graph of every decision, commitment, and deadline. All stored locally as Markdown files you can edit, like Obsidian but with an AI that actually uses the connections.
The demo cases are telling:
- "Build me a deck about our next quarter roadmap" pulls actual priorities and commitments from your conversation history
- "Prep me for my meeting with [person]" surfaces past decisions and open questions you've forgotten
- Voice memos automatically update the graph when plans change, linking back to original commitments
This isn't just another RAG system with file uploads. The founders' previous startup was acquired by Coinbase specifically for their graph neural network expertise—they understand how relationships between information create intelligence.
What Nobody Is Talking About
While everyone's obsessing over which LLM is smartest, Rowboat tackles the real bottleneck: institutional memory. Most companies lose context every time someone leaves, every time a project shifts, every time a decision gets made in Slack and forgotten.
The technical architecture is surprisingly elegant. Two components: a living context graph that auto-updates from your communications, and a local assistant agent with shell access and Model Context Protocol support. Apache-2.0 licensed, works with any LLM including local models, stores everything as readable Markdown.
But here's what's actually fascinating: this solves the "AI agents are impressive demos but useless in practice" problem. Most agents fail because they lack context about what matters, what's been tried, what's been decided.
Rowboat's positioning around "enterprise multi-agent AI adoption" suggests they're thinking bigger than personal productivity. Their examples span fintech, insurance, telecom—sectors where institutional memory and complex decision trees actually matter.
The codebase is 96.4% TypeScript with active development (211 closed PRs), supporting custom LLM providers, multi-agent orchestration, and API integration. Downloads available for Mac, Windows, Linux.
The Uncomfortable Truth
This feels like the first AI tool that might actually change how knowledge work happens, not because it's smarter, but because it remembers. When your AI assistant knows that the marketing budget discussion is connected to the hiring freeze which relates to the product roadmap delay, suddenly "build me a realistic Q2 plan" becomes possible.
The open-source model with optional cloud hosting is smart positioning—developers can experiment locally while enterprises get hosted solutions.
The real question isn't whether Rowboat will work. It's whether you're comfortable with an AI that has perfect recall of every commitment you've made, every deadline you've missed, and every priority that's shifted without explanation.
Because once you have an AI coworker with a better memory than yours, there's no going back to pretending you don't remember what you promised last month.
