> ## Content Index
> Fetch the complete content index at: https://nexi.fund/llms.txt
> Use this file to discover other available public pages before exploring further.

# Graphon Labs Raises $35M to Sell Memory as AI Infrastructure
- URL: https://nexi.fund/graphon-labs-agent-memory-2026/
- Published: 2026-11-11T09:30:08.000Z
- Updated: 2026-11-11T09:30:08.000Z
- Description: Graphon Labs has raised $43.3M in seven months to sell a model-neutral memory layer for AI agents. Its benchmarks lead the field, but they are vendor-run — and independent evidence says memory buys cost and latency, not accuracy.
- Author: Nexi.fund Labs
- Tags: AI & Infrastructure, #mode-6, #hook-paradox, #track-A, #layout-signal-grid

Graphon Labs announced a $35 million Series A on November 10, led by RRE Ventures, with HP, Energy Impact Partners, Hitachi Ventures and Gaia VC participating. The round brings total funding to $43.3 million, six months after an $8.3 million seed took the company out of stealth.

Twenty people. Sixteen months old. A product that files what an AI agent has already seen.

That last part is the whole bet. Graphon sells memory: a layer that ingests video, images, audio and documents, builds a graph of how the pieces connect, and keeps track of when each fact was true and which source it came from. Agents, in this telling, fail less because they cannot reason and more because they cannot remember.

The claim runs against the direction of the model vendors. Context windows keep growing, and the big assistants now ship user-level memory by default. Graphon's answer is a layer that sits beneath Claude, GPT or anything else in production and stays vendor-neutral.

## A $35M bet that memory, not context, is the constraint

Arbaaz Khan, the chief executive, compares vector search to searching through the pages of a torn-up book: the words are all there, the order and the relationships are not. Retrieval-augmented generation returns chunks. It does not return the fact that the refund policy changed in March, or that the 70V and 770V entries in a maintenance log refer to different machines.

Graphon's numbers are specific. On LoCoMo, a long-conversation memory test, the company reports 93.9% with 58% fewer tokens than loading the whole conversation into the model. On ConvoMem it reports 90.2% at 95% fewer tokens. Both comparisons use baselines published by Databricks.

What those figures prove, and what they do not, runs through the rest of this piece. So does the architecture underneath them, the competing memory layers from Mem0, Zep, Letta and Oracle, and the one risk that decides whether memory becomes infrastructure or a checkbox.

$43.3M Total raised 

#### Graphon Labs cumulative funding

$8.3M seed in May 2026, followed by a $35M Series A in November, both inside seven months of stealth exit. · *Business Wire, 2026*

## What the graphon actually does

The name is not branding. A graphon is a mathematical object used to approximate very large graphs, the kind with too many nodes to analyze directly. Khan and co-founder Clark Zhang worked on graph-based memory for robots during PhDs at the University of Pennsylvania's GRASP Lab. The company describes its approach as the third generation, after vector RAG and graph RAG.

The architecture has two moving parts. One is a detailed graph that stores information and its relationships. The other is a compressed graphon representation that guides retrieval through that larger structure. Ingested video, audio, images and text are turned into an auto-generated ontology, with links back to the original source.

Time is treated as a first-class field. A record carries what was known, when it became true, and which source established it. For a customer-service agent, that is the difference between a current contract and last quarter's version. For an analyst, it is the difference between a fact and a stale excerpt that happens to rank well on similarity.

For developers, the interface borrows from object storage: buckets, objects, files, queries. Keyword, semantic and image search sit on top, and an agent can ask questions across several sources. The point is to avoid rebuilding a retrieval stack for each data type and each new model.

The product line is three-deep. Graphon Enterprise runs on Graphon's cloud, a Graphon-managed or customer-managed VPC, or on-premises. Graphon Developer is for builders. Graphon On-Device Memory targets phones, laptops, cameras and wearables. The company holds SOC 2 Type II, and its core technology is patent-pending.

> Models can reason, and agents can act, but neither gets better if they cannot build off of prior knowledge.— Raju Rishi, General Partner, RRE Ventures

The customer list needs reading with care. Graphon says it is "already deployed with the San Francisco 49ers, Hitachi, and other large enterprises." The 49ers' executive vice president of technology, Costa Kladianos, describes exploring uses such as finding sponsorship opportunities. Hitachi Rail's Amaury Jourdan describes an evaluation that "shows promise" in an industrial setting. HP is named as a user. The South Korean conglomerate GS signed on as an early customer during the seed round, with convenience-store footfall analysis and CCTV construction-safety work.

Evaluations are not contracts. That distinction matters at a $35 million round, and it is the first thing to watch over the next two quarters.

## The leaderboard is mostly self-reported

Graphon publishes results on seven benchmarks. The LongMemEval table is the most useful, because it lines up the whole field on one test of long-term chat memory.

| System                 | LongMemEval accuracy | Basis                                           |
| ---------------------- | -------------------- | ----------------------------------------------- |
| **Graphon Labs**       | 96.2% (481/500)      | Vendor run                                      |
| Agent Zero Memory      | 95.6%                | Research paper                                  |
| Mem0                   | 94.4%                | Vendor-reported; 73.8% on an independent re-run |
| Oracle AI Agent Memory | 93.8% (469/500)      | Vendor evaluation                               |
| Zep                    | 71.2%                | Reported by a third party                       |

LongMemEval figures as compiled by Graphon Labs, November 2026\. Vendor scores are not independently audited.

Two things stand out. Almost every number on the board is reported by the company that benefits from it. And the one visible independent re-run, Mem0 at 73.8%, lands more than 20 points below Mem0's own 94.4%.

Graphon is not immune to the pattern. The press release sells 93.9% on LoCoMo. The company's own benchmark page reports 94.0% at 7.9K tokens per question, 92.8% on LongMemEval at 15.1K tokens, and a best-case 96.2%. The spread inside one company's materials is a caution, not a scandal. Benchmarks in this category are young enough that a single configuration choice can move a score by double digits.

The structural problem is that memory systems are tested against each other on tasks that reward different things. LoCoMo rewards recall across long conversations. BEAM pushes to ten million tokens and tests contradictions, changing facts and event order. DolphinBench scores task success and reports cost and latency beside it. A vendor can top one chart and lose on another without either result being false.

## Memory is a cost technology before it is an accuracy technology

The independent evidence points somewhere less flattering. A March 2026 study compared a Mem0-style fact layer with long-context inference on GPT-5-mini across LongMemEval, LoCoMo and PersonaMem v2\. On two of the three benchmarks, full context was roughly 33 to 35 percentage points more accurate. The memory layer matched it only on PersonaMem v2, where consistent persona traits matter.

Where the memory layer wins is the bill. Full context pays per turn, and the cost grows with the history. A memory layer pays a one-time write and then nearly constant read costs. At 100,000 tokens of history, break-even lands near ten interactions; after 20 questions the saving is around 26%. The longer the history, the earlier the crossover.

As we wrote in [November](https://nexi.fund/agentic-ai-cost-per-task/), the token bill is where agent budgets break. A memory layer is one more way to cap it, and the same logic applies to latency: Oracle reports that its AI Agent Memory held input near 1,300 tokens per request in an 80-turn conversation, against more than 13,900 for a flat-history baseline.

The research frontier is less settled than the marketing suggests. LongMemEval-V2, published in May 2026, tests whether a memory system can give an agent the working knowledge of an experienced colleague: 451 questions across environments with up to 500 trajectories and 115 million tokens. The best method reported, AgentRunbook-C, reached 72.5%, against 48.5% for the strongest RAG baseline and 69.3% for an off-the-shelf coding agent. Its authors note the latency cost. The accuracy-latency frontier is still moving.

81.7% ATM-Bench QA score 

#### Multimodal memory benchmark

Graphon's reported score, 5.0 points above the next system on the public leaderboard. · *Graphon Labs, 2026*

## Who else is selling memory

The field is more crowded than a single leaderboard implies. Mem0 extracts facts with an add-and-search API over a hybrid vector and graph store, and its paper claims 91% lower p95 latency with more than 90% token savings. Zep runs a temporal knowledge graph whose edges carry validity windows, so a fact can expire instead of lingering. Letta, descended from MemGPT, moves memory into the runtime and lets the agent page its own context. Supermemory, Databricks' MemCurrent and Oracle's database-native layer fill in the rest.

💡

**Three architectures, three bets**  
  
Extraction-first (Mem0): cheapest to integrate, weakest on relationships.  
  
Temporal graph (Zep, Graphon): best on facts that change, heaviest to operate.  
  
Runtime paging (Letta): strongest for long-running agents, but you adopt a runtime.  
  
The differentiator is moving from data structure to deployment surface. 

The harder question is whether per-agent memory composes into company memory. A layer that remembers one agent's sessions solves a demo. An organisation needs one version of a fact, with permissioning, across dozens of agents and several tools. That is a harder product than a retrieval API, and it is where the enterprise budget sits.

## The bundling risk

Model vendors are not standing still. Assistants ship user-level memory, context windows are large enough that "put everything in the prompt" survives many workflows, and every hyperscaler sells a managed retrieval product. The independent study above is, in effect, the strongest argument for bundling: accuracy is better when you keep everything.

Graphon's counters are governance, neutrality and cost. Provenance and temporal validity are audit features that a chat memory feature does not provide. SOC 2 Type II answers a procurement question that a consumer toggle does not. On-device memory is a market the model labs have not entered.

The 49ers use case shows the wedge clearly. A question that spans contracts, video and years of sponsorship history is expensive to answer with a giant prompt and awkward to answer with snippet retrieval.

Khan's framing is that this is "software as memory, where what you keep is understood, not just stored." The sentence is good marketing. It is also a testable product claim: either the retained knowledge improves an agent's output on a real task, or it is a cheaper index with a better story.

The risk is absorption. If org-level memory with provenance arrives inside the next model release, a standalone layer has to justify itself on price and portability alone. That is a real business. It is a smaller one than "the source of truth beneath every agent."

🎯

The memory layer is being priced as infrastructure: $35M for a 20-person team whose public proof is evaluations, not contracts.  
  
The numbers that sell it are vendor-run; the independent data says memory buys cost and latency, not accuracy.  
  
The real test is whether any independent memory layer survives the next model release. 

📊

**Signals to watch**  
  
Converted enterprise contracts at HP, Hitachi or the 49ers by the second quarter of 2027.  
  
An independent Graphon run on the open Agent Memory Benchmark harness.  
  
Model vendors shipping organisation-level memory with provenance and audit trails.  
  
Whether memory is priced per seat or per token — the tell for infrastructure versus feature. 

[ Graphon Labs raises $35M Series A to accelerate AGI with Persistent Intelligence The primary announcement: round size, lead investor, participants, product line and the benchmark claims the company chose to lead with. Business Wire ](https://www.businesswire.com/news/home/20261110959095/en/Graphon-Labs-Raises-%2435M-Series-A-to-Accelerate-AGI-with-Persistent-Intelligence?ref=nexi.fund) 

The source for every funding figure and the deployment language that later coverage repeated.

[ Graphon Labs — persistent intelligence, architecture and benchmark board The company's own technical account: graphons, the two-part architecture, the three products and the full LongMemEval comparison table. Graphon Labs ](https://www.graphonlabs.ai/about?ref=nexi.fund) 

Useful precisely because it lists rivals' scores beside its own, including the Mem0 gap between claimed and independently re-run results.

[ Graphon Labs raises $35m to give AI agents a memory that lasts Independent reporting on the round, the 20-person team, the token savings on LoCoMo and the customer names behind the deployment claim. The Next Web ](https://thenextweb.com/news/graphon-labs-35m-series-a-ai-agent-memory?ref=nexi.fund) 

Clarifies that the 49ers and Hitachi Rail are exploring and evaluating, which is narrower than "deployed".

[ A practitioner comparison of Mem0, Zep and Letta, the benchmark canon, and the March 2026 cost study that found full context more accurate on two of three tests. Agent memory 2026: why the memory layer is not a benchmark victory saaro ](https://blog.saaro.net/en/agent-memory-2026-warum-die-gedaechtnisschicht-kein-benchmar?ref=nexi.fund) 

The clearest independent statement of the category's uncomfortable finding: memory is a cost and latency technology first.

[ LongMemEval-V2: evaluating long-term agent memory toward experienced colleagues The benchmark that moves past chat recall: 451 questions over environments with up to 500 trajectories and 115M tokens, plus the accuracy-latency trade-off in current methods. arXiv ](https://arxiv.org/abs/2605.12493?ref=nexi.fund) 

Sets the ceiling for the category: the best reported method reaches 72.5%, not the 90s the vendor charts suggest.