Graphon Raises $35M to Give AI Agents Memory Beyond the Context Window

San Francisco-based Graphon Labs raised a $35 million Series A led by RRE Ventures on November 10, 2026, bringing its total funding to $43.3 million less than six months after an $8.3 million seed round. HP, Energy Impact Partners, Hitachi Ventures and Gaia VC also participated.

The financing backs a persistent-memory platform intended to help AI agents recover earlier information after it has left a model’s context window. Graphon’s proposition is that relationships among sources should be maintained as reusable infrastructure, rather than reconstructed by each application when it needs an answer. The funding is established; the performance case rests on tests with different levels of outside review.

What the Series A will support

The near-term spending plan centers on people who can build and sell the system. The Caplight funding record lists engineers and researchers, along with sales staff, in California as intended hires. The roles reflect both sides of Graphon’s challenge: maintaining a complex retrieval system as information changes and finding organizations willing to use it across their applications.

That commercial work is consequential for an infrastructure startup. A memory layer has to fit into existing storage, search and agent workflows before its technical design can become a recurring service. The round gives Graphon room to pursue deployments, but investment in the thesis is distinct from evidence that its approach will be cheaper or more reliable in routine use.

Why Graphon wants retrieval to keep relationships

Conventional retrieval-augmented generation, or RAG, typically divides material into passages, retrieves likely matches and places them in a model’s prompt. It can give an agent access to updated information without retraining the model. The answer still depends on whether the retrieved passages include the context that changes a fact’s meaning.

Consider a refund policy split between a general rule and an exception for sale items. A search might return the rule because its wording closely matches a customer’s question while missing the exception elsewhere. A well-tuned RAG stack can reduce that risk with exact matching, metadata, reranking or additional passages, but those components need to be configured and maintained as the source material changes.

Graphon’s design instead represents details and their connections in a graph, then uses a compressed representation called a graphon to guide retrieval through it. A video event can remain associated with a spoken explanation; a spreadsheet value can retain the headings needed to interpret it; a document rule can stay linked to its exception. The proposed benefit is a path from an answer back through the relevant relationships to its sources, even when the original material exceeds the agent’s immediate context.

For developers, the platform presents storage and search concepts such as buckets, objects, files and queries. Keyword, semantic and image searches can operate alongside questions that require connections across several sources. That shared interface is central to the business thesis: if it works across different kinds of information, applications may need fewer separately assembled retrieval components. Its value will depend on how well it handles changing data, source tracing, speed and cost in deployment.

What the benchmark claims actually measure

In Graphon’s launch essay, founder and chief executive Arbaaz Khan asks, “why are we throwing away useful structure in the first place?” and reports that a Graphon plus GPT 5.6 Luna configuration passed 504 of 600 DolphinBench tasks, compared with 424 for the best public leaderboard run using the same answer model. Graphon ran its own configuration and submitted the result for review; the comparison run was already on the public leaderboard.

The company lists a total cost of $11.84 for its run, including indexing, search and the answer model’s tokens, against $96.21 for that public comparator. Those figures describe complete configurations under the benchmark’s conditions. They do not isolate the contribution of the graph representation, and a company-run submission should retain that label until the leaderboard reviews it.

The multimodal results have a separate status distinction. Graphon appears on the public ATM-Bench leaderboard, while its ATM-Bench Hard submission was awaiting review at launch. Its conversational-memory comparisons also combine Graphon’s own runs with previously published baselines, rather than a fresh independent evaluation of every system under one operator. These results indicate the kinds of recall and cross-source questions the platform targets, with the review status determining how much weight each ranking can carry.

The test of a standalone memory layer

Graphon is trying to make persistent memory a service beneath agents that work with growing collections of documents, conversations, images, audio and video. A larger context window can hold more material during a task, but it does not by itself select which earlier evidence matters or track how a fact changed across sources. Graphon’s graph-based design is an attempt to perform that selection repeatedly without making every application build its own retrieval stack.

The next evidence will come from reviewed benchmark submissions and sustained deployments. For organizations using agents over changing information, the deciding question is whether Graphon can preserve useful connections while keeping retrieval fast, traceable and economical as the collection grows.

Also read:

• Suki Raises $90M—Its Next AI Agents Move Beyond Clinical Notes
• Hadrian Raises $40M—Offensive AI Becomes a Funding Category
• Vocca Raises $20M—One Million Monthly AI Calls Still Need Context