Graphon
Category: AI Infrastructure
Graphon Labs builds a persistent-memory and intelligence layer for AI agents. It uses a graph-based representation (graphons) to connect multimodal data across documents, video, audio and images. Graphon was founded in 2025. The company is led by Arbaaz Khan. Based in San Francisco, United States. Team size: 11-50. Total funding raised: $43.3M. Latest round: Series A. Key investors include RRE Ventures, HP, Energy Impact Partners, Hitachi Ventures, Gaia VC, Novera Ventures, Perplexity Fund, Samsung Next, GS Futures, B37 Ventures, Aurum Partners.
- Founded
- 2025
- Headquarters
- San Francisco, United States
- Team size
- 11-50
- Total funding
- $43.3M
Value proposition
Persistent intelligence for AI agents. Graphon builds a global, inferred graph over a whole corpus so that agents keep relationships, history and sources instead of losing them to chunked RAG retrieval.
Products and solutions
Graphon Enterprise, Graphon developer platform (buckets, objects, files and queries, with keyword, semantic and image search plus cross-source questions), on-device memory for phones, laptops, cameras and wearables.
Unique value
Described as the third generation after vector RAG and graph RAG. It scores each piece of data by its position in the whole corpus using an inferred global schema, so facts keep their relations and sources.
Target customer
Enterprises with large multimodal data estates (manufacturing, CPG, financial services, sports, media and entertainment), plus developers building AI agents.
Industries served
Manufacturing and rail, Consumer Packaged Goods, Financial Services, Sports, Media and Entertainment, Retail and construction safety (GS Group).
Technology advantage
Graphon functions, a mathematical framework for representing very large graphs, developed from the founders' PhD research at Penn's GRASP Lab. Technical advisors Jennifer Chayes and Christian Borgs of UC Berkeley co-formalised the underlying graphon concept. Self-reported benchmark results include LoCoMo 93.9-94.0%, ConvoMem 91.0%, LongMemEval 92.8%, ATM-Bench (#1/#2) and DolphinBench 504/600. Some submissions (ATM-Bench Hard, DolphinBench) were awaiting review at launch.
How they differentiate
Multimodal (video, audio, image, text) memory built on a graph structure inferred across the whole corpus, with source tracing. Mem0, Zep and Letta focus mainly on text and conversational memory with local schemas. Graphon reports lower token cost per question than full-context approaches.
Main competitors
Mem0, Zep (Graphiti), Letta
Key partnerships
HP, Hitachi (including Hitachi Rail evaluation), GS Group (Korea), San Francisco 49ers, technical advisors Jennifer Chayes and Christian Borgs (UC Berkeley).
Notable customers
San Francisco 49ers, HP, Hitachi (Hitachi Rail evaluating), GS Group
Major milestones
Emerged from stealth with an $8.3M seed on 2026-05-14. Raised a $35M Series A led by RRE Ventures on 2026-11-10. Reports top results on LoCoMo, ConvoMem, LongMemEval and ATM-Bench. Launched three products: Enterprise, developer platform and on-device memory.
Growth metrics
About 20 employees as of November 2026. Raised $43.3M in total within about six months of leaving stealth.
Market positioning
Standalone persistent-memory infrastructure layer beneath AI agents, with an enterprise-first go-to-market and later expansion to devices and consumer products.
Geographic focus
US-led, with enterprise customers and investors in the US, Korea and Japan.
Patents and IP
Patent-pending multimodal ingestion technology (company-stated).
About Arbaaz Khan
Ex-Amazon Senior Applied Scientist (5+ years); Robotics MSE 2018 and PhD in ESE 2021, University of Pennsylvania GRASP Lab. Built graph-based memory for robots during his PhD with co-founder Clark Zhang.
Latest news about Graphon
More AI Infrastructure companies
Official website: https://graphon.ai