Graphon
by Graphon Labs
Graph-based persistent memory for enterprise AI agents across documents, video, images and logs
Graphon is a memory layer for enterprise AI agents. It turns documents, video, images and logs into a persistent knowledge graph that any model can draw on. It is aimed at enterprise AI and data teams whose agents lose relationships and timing when they rely on context windows and plain RAG retrieval.
Graphon, built by San Francisco startup Graphon Labs, is what the company calls a pre-model intelligence layer: software that prepares an organisation's data before a foundation model sees it, instead of stuffing raw text into the context window. It ingests documents, video, audio, images, logs and structured databases, then builds a self-supervised, schema-free knowledge graph that records how pieces of information connect, where each came from and when it changed. The approach rests on graphons, a mathematical framework for very large graphs; UC Berkeley professor Christian Borgs, who helped develop graphon theory, is a technical adviser, and SiliconANGLE reported that the platform uses small models of roughly 200 million parameters to process data as graphs. Graphon is model-agnostic, so teams can point Claude, GPT or other models and agent frameworks at it. The company came out of stealth on May 14, 2026 with an $8.3 million seed led by Novera Ventures, then closed a $35 million Series A led by RRE Ventures on November 10, 2026, bringing total funding to $43.3 million. GS Group in South Korea runs it in production for retail customer-movement analysis and CCTV safety analytics on construction sites; HP, Hitachi Rail and the San Francisco 49ers are evaluating it. Graphon reports 93.9% on the LoCoMo long-conversation memory benchmark using 58% fewer tokens than loading the full conversation, and 81.7% on Cambridge's ATM-Bench. Both figures are company-reported. It competes with agent-memory tools such as Zep, Mem0 and Letta and with GraphRAG-style retrieval stacks.
An enterprise AI platform lead whose agents must reason over large multimodal archives (video, documents, sensor and log data) where plain RAG misses how records relate over time.
Agents get a persistent, provenance-tracked memory of how the organisation's data connects, at lower token cost than loading full context.
At a Glance
- Category
- Enterprise Search & Knowledge
- Pricing
- Contact for pricing
- Target Market
- CTOs, Heads of AI, Enterprise Developers, Data Scientists
- Deployment
- API-based
- Founded
- 2025
- Headquarters
- San Francisco, United States
- Team Size
- 11-50
Key Features
- ✓Multimodal ingestion
Ingests documents, video, audio, images, logs and structured databases into one representation, so agents can reason across formats instead of text alone.
- ✓Schema-free knowledge graph
Builds a self-supervised graph of how records relate without a hand-designed schema, which removes the ontology work most graph projects stall on.
- ✓Provenance and change tracking
Records where each piece of information came from and when it changed, giving agents a timeline and giving reviewers a trail to check answers against.
- ✓Model-agnostic memory layer
Works beneath any foundation model or agent framework, so teams can point Claude, GPT or open models at the same memory without rebuilding it.
- ✓Token-efficient recall
Serves structured memory instead of full raw context; the company reports 58% fewer tokens than full-conversation loading on the LoCoMo benchmark.
- ✓On-device memory roadmap
The company plans memory for phones, laptops, cameras and wearables after its enterprise focus, relevant for edge and camera-heavy deployments.
Capabilities
Use Cases
- •Construction site safety analytics
GS Group links CCTV footage with site data in real time to surface safety risks on industrial construction sites.
- •Retail customer-movement analysis
Convenience-store operators map how customers move through stores by connecting camera data with store records, informing layout and staffing decisions.
- •Enterprise content management
Organisations connect video, audio, image and document archives so agents can answer questions that span formats and track how facts changed.
- •Industrial process and compliance review
Industrial teams such as Hitachi Rail evaluate it to find process gaps, safety risks and compliance issues hidden across logs and documents.
- •Commercial opportunity discovery
The San Francisco 49ers are exploring it to find sponsorship opportunities by relating partner, audience and business data.
Ideal For
Best For
- ✓Industrial and physical-operations data where CCTV video, logs and documents must be linked to spot safety or process issues
- ✓Agent workflows that need long-term memory across sessions rather than one context window
- ✓Retail analytics teams combining camera footage with store and customer data
- ✓Enterprises that want one memory layer usable by several foundation models and agent frameworks
- ✓Relationship-heavy analysis, such as telling whether separate signals belong to one event or campaign
Not Ideal For
- ✗Teams that need published pricing, self-serve sign-up and public documentation today; Graphon discloses none of these
- ✗Buyers who require independent benchmark validation or multiple production references, since current results are company-reported and most named users are still evaluating
- ✗Simple FAQ or single-document question answering, where a standard RAG pipeline is cheaper and adequate
Deployment
Market Analysis
Pros
- ✓Strong investor and corporate backing: $43.3M raised in six months, with HP and Hitachi investing as well as evaluating
- ✓One production deployment at GS Group across retail and industrial safety
- ✓Multimodal and model-agnostic design avoids lock-in to one model vendor
- ✓Reported LoCoMo result pairs high accuracy with a 58% token reduction
Cons
- ✗Benchmark results (LoCoMo 93.9%, ATM-Bench 81.7%) are company-reported; none of the coverage cites independent testing
- ✗Early-stage: about 20 employees, one named production customer, and HP, Hitachi Rail and the 49ers still exploring or evaluating
- ✗No public pricing, documentation, SDK or compliance certifications were found, which slows procurement review
Pricing
Graphon Enterprise
Contact for pricing
- ✓Persistent multimodal memory layer
- ✓Model-agnostic integration
- ✓Enterprise deployments
Graphon has published no list pricing, metering basis or plan tiers as of November 2026. Coverage of both funding rounds omits pricing, so expect a sales-led enterprise contract.
Security & Compliance
Sources
This page was written from 5 sources, 5 on domains other than graphonlabs.ai.
- 1.thenextweb.com — graphon labs 35m series a ai agent memory
- 2.cryptobriefing.com — graphon raises 35m ai agent memory
- 3.vktr.com — graphon ai raises 83m to build premodel intelligence layer
- 4.siliconangle.com — siliconangle.com
- 5.aiweekly.co — graphon labs raises 35m series a from rre ventures to give a
Stay Ahead of the Curve
Weekly enterprise AI insights for technology leaders. No spam, no vendor pitches—unsubscribe anytime.
SubscribeRelated Products
Otter.ai
AI meeting notetaker turned company-wide meeting knowledge base, with an API, MCP server and agentic chat
LightOn Suite
European enterprise document AI that turns company files into a governed knowledge base and the apps built on it, hosted in France
Atlassian Rovo
Atlassian's AI search, chat and agent layer across Jira, Confluence and connected apps, metered in pooled Rovo credits
Vectara
Enterprise RAG and agent platform with built-in hallucination scoring for regulated industries