the problem
Modern AI models are powerful, but still constrained by limited context.
Enterprises store trillions of tokens across documents, videos, logs, databases, meetings, and internal systems. Traditional retrieval systems like RAG can surface relevant snippets, but they cannot understand how information connects across datasets, modalities, and time.
As a result, AI agents and agentic systems can answer questions about isolated pieces of data, but struggle to reason across the full organizational context needed for accuracy and reliability.
what graphon is
Graphon is a persistent intelligence platform for multimodal data. It ingests videos, images, audio, and documents using patent-pending technology, goes deep into the data to automatically build an ontology, and gives AI the context it needs.
Because it goes that deep, you are not limited to searching for broad entities like “blue shirts.” You could ask for “blue shirts with five buttons, two pockets, and short sleeves,” and see exactly where each result came from.
The technology rests on graphons, a mathematical function for representing very large graphs. The founders developed the underlying research during their PhDs at Penn's GRASP Lab.
how we're different
generation one
Scores each piece of data alone. It hands back a list of look-alikes and stops there.
what gets lost
Relations, history and order. Similar-looking pieces crowd out the real answer.
generation two
Scores a chunk plus its neighbors. Still a candidate list, built inside the radius you explored.
what gets lost
Anything humans didn’t map. Anything outside the radius, or that the schema didn’t foresee.
generation three
Scores each piece of data by where it sits in the whole corpus. Search, analyze and find directly against the structure.
what gets lost
Nothing. Every fact keeps its place, its relations and its source.
the benchmarks
Graphon beats every existing memory system on accuracy and on cost, while being truly multimodal across video, audio and text. It sets a new cost-accuracy Pareto frontier.
The first benchmark for multimodal, multi-source personalized referential memory QA over long time horizons (~4 years), with evidence-grounded retrieval and answering.
Graphon is #1 and #2 on the leaderboard.
A benchmark for very long-term conversational memory: dialogues of about 600 turns across up to 32 sessions, with questions about what was said, when, and how it connects.
Graphon redraws the cost-accuracy Pareto frontier: 93.9% accuracy at 7.9K tokens per question, more accurate than full context with less than half the tokens.
A benchmark of 75,336 question-answer pairs over long conversation histories, testing user facts, assistant recall, preferences, facts that change, implicit connections, and knowing when not to answer.
Graphon redraws the cost-accuracy Pareto frontier: 90.2% accuracy at 8.9K tokens per question, more accurate than full context with a twentieth of the tokens.
A benchmark for agent memory in action: three simulated knowledge workers with years of history each, and 600 tool-using tasks that pass only if the agent recalls and applies the right fact. Every run reports its total cost and latency alongside accuracy.
Graphon passes 504 of 600 tasks (84.0%) for $11.84 in total, with a median of 28.8 s a task. The best run on the public leaderboard, on the same model, passes 424 (70.7%) for $96.21.
That $11.84 is the whole bill for the run: Graphon’s indexing and search, plus every input and output token GPT 5.6 Luna spent working through the 600 tasks.
| Memory | Harness + model | Passed | Accuracy | Total cost | Median latency |
|---|
team and vision
leadership
Graphon AI was founded by researchers and engineers with deep expertise in AI, machine learning systems, and large-scale enterprise infrastructure. The leadership team includes former researchers and engineers from Amazon, Meta, Google, Apple, PayPal, NVIDIA, Rivian, Samsung AI Center, NASA, and leading research institutions.
founders
Founder & CEO
5+ years as Senior Applied Scientist, serving 10M+ customers. PhD and Masters in ML and Robotics.
Co-Founder & CTO
5+ years as Senior Machine Learning Engineer. PhD and Masters in ML and Robotics.
advisors
technical advisors
Jennifer Chayes
Dean, College of Computing, Data Science, and Society, UC Berkeley
Christian Borgs
Professor of Computer Science, UC Berkeley, Co-creator of graphon theory
Alejandro Ribeiro
Professor, University of Pennsylvania ESE Department
business advisors
Bask Iyer
Ex-CIO of VMWare, Honeywell, and Juniper Networks
Alan Fletcher
Ex-COO of Regrello (acquired by Salesforce)
Simon Mulcahy
Ex-Chief Innovation Officer of Salesforce
Akash Bhatia
Senior Partner and Global Tech Sector Lead at BCG
the team
key facts
questions
No. RAG retrieves snippets that look relevant to a question. Graphon builds a persistent, relational memory of how your content connects, across modalities and over time, so AI can reason across all of it instead of over isolated pieces.
Video, images, audio and documents. It reads the content itself, the frames, the speech and the pages, rather than the file names and labels around it, and every answer shows where it came from.
Graphon builds the ontology automatically as it reads, down to details a team would not model by hand. So you can ask for "blue shirts with five buttons, two pockets and short sleeves," not just "blue shirts."
Wherever it is allowed to. Graphon runs in our cloud, inside your own VPC, or on a device next to where the data is made, with nothing leaving the building.
Graphon is SOC 2 Type II certified. The live control status, and the reports on request, are on the Graphon Trust Center.
You can start free, with no card needed, or book a demo and the team will walk you through it.
careers
Be part of building the next generation of AI capabilities beyond today's models.
View Open Roles →brand assets
Writing about us or putting us on a slide? The logo is ready to download in black, white and cyan, as SVG, PNG and JPEG.
Get the logos →