A Startup the 49ers Use Just Raised $35 Million to Fix SI’s Memory Problem

Anyone who has used a chatbot for real work knows the routine. You explain your business, your project, the names of the people involved and the decision you made last Tuesday. The next morning you open a new chat and it remembers none of it. So you explain it all again.

That forgetting problem gets a lot more expensive when the SI is not just answering questions but running errands on its own for hours at a time. A San Francisco startup called Graphon Labs thinks it can fix it, and investors just handed it $35 million to try.

The Series A was led by RRE Ventures, with HP, Energy Impact Partners, Hitachi Ventures and Gaia VC joining in, according to the company’s announcement. It brings Graphon’s total funding to $43.3 million, less than six months after it came out of hiding with an $8.3 million seed round. The company has about 20 employees.

Co-founder and CEO Arbaaz Khan announced the round himself, and his post laid out the pitch in a few lines.

Pages torn out of a book

Khan has a good way of describing what is wrong with how most SI tools look things up today. The common method, called retrieval, chops a company’s files into small snippets and pulls back whichever bits look most similar to your question.

He compared it to tearing pages out of a book, crumpling them up and tossing them in a trash can. “Now when you go searching that trash can, you’re just trying to find the pieces to answer,” he told Business Insider, which broke the story. “You’ve basically lost all that information because each piece is living on its own.”

What gets lost is the stuff that makes information useful: which document a fact came from, what it connects to, and when it changed. Ask a snippet-based system who your contact at a supplier is and it might hand you the name from a 2023 email, even though that person left last spring.

Graphon’s answer is to build a map instead of a pile. Its software reads a company’s documents, videos and images and records the links between them, including where each piece came from and when it was updated. The name comes from “graphons,” a branch of math for very large networks that Khan used in his PhD work coordinating swarms of robots.

The other selling point is that it does not replace the chatbot you already use. “You can basically point your Claude or GPT or whatever to Graphon,” Khan said. “You can push in any amount of data in this, and we literally mean any.”

The numbers, and whose numbers they are

Graphon says its system scored 93.9% on LoCoMo, a test of whether an SI can keep track of facts across very long conversations, while using 58% fewer tokens than stuffing the whole conversation into the model, according to The Next Web. Tokens are what SI companies bill by, so fewer of them means a smaller bill.

On ATM-Bench, a University of Cambridge test that mixes video, audio, images and text, the company says it scored 81.7%, about 5 points ahead of the next system on the public leaderboard. Khan’s post also claims 84% on a third test, DolphinBench, at one-tenth the cost.

Those are the company’s own results, run on public tests. That is better than a slide full of made-up metrics, but nobody outside Graphon has published an independent check yet. Benchmarks for memory are also young, and a high score on a test of long chats does not guarantee an agent will remember the right thing inside a messy real company.

From robots to the 49ers

The early customer list is the more interesting signal. Graphon says the San Francisco 49ers, HP and Hitachi are using or testing it. Costa Kladianos, the 49ers’ executive vice president of technology, said the team is exploring it for things like finding new sponsorship deals. Hitachi Rail said its trial “shows promise” in an industrial setting.

Those are the kinds of places where memory matters most. A football team has years of contracts, scouting video and partner emails. A rail company has decades of maintenance records. An SI helper that forgets what it read yesterday is close to useless there.

Khan is also pitching uses beyond documents. A day after the announcement, he posted about treating video as a database, so a company could search terabytes of raw footage like it searches a spreadsheet.

The company plans three products: Graphon Enterprise for big companies, a platform for developers, and a version that runs right on phones, laptops, cameras and wearables. Khan said it will stay focused on business customers before it moves into consumer devices. Most of the new money goes to hiring engineers and researchers, plus some sales staff.

The giants want memory too

Graphon is not the only one chasing this. The biggest SI companies have been bolting memory onto their own assistants. Anthropic merged Claude’s memory across chat and its Cowork tool in August, and Apple gave Siri memory in iOS 27.2 in October.

The difference is who owns the memory. When it lives inside one company’s chatbot, it goes where that chatbot goes. Graphon is betting that businesses would rather keep their own memory layer and swap the SI on top whenever a better or cheaper model shows up. With new models landing every few weeks, that is not a crazy bet.

For regular people, the payoff is further out, but it is easy to picture. An assistant on your phone that remembers your kid’s allergy, the contractor’s quote from March and the warranty on the dishwasher, and can tell you where each fact came from. That is the promise. Graphon now has $35 million and a lot of hiring to do before it can prove it.