In 2001, Chenggong Charles Fan submitted a doctoral dissertation at Caltech about keeping clusters of networking equipment alive when one part failed. Twenty-five years later, his name appeared on a paper about helping personalized AI agents keep their memories intact. Between those documents came two startups, an acquisition, an R&D center in China, a storage business inside VMware and a long campaign to persuade computers to stop treating memory as something small, local and disposable.
The cleanest way to understand Fan's career is not as a march from one fashionable technology to another. It is as one systems question asked at increasing scale: how can useful state survive when the machines beneath it change?
The answer first involved network services. Then it involved files and virtual machines. At MemVerge, it involved persistent memory, cloud workloads and shared memory connected over CXL. Now it involves the context an AI agent needs to recognize a user from one session to the next. The nouns keep changing. The concern is continuity.
Chapter one
A thesis with a commercial afterlife
Fan studied electrical engineering at Cooper Union, graduating summa cum laude in 1995. He went west for graduate work at Caltech, completing a master's degree in 1996 and a Ph.D. in 2001. His dissertation, “Fault-tolerant cluster of networking elements,” joined distributed systems and computer networks. The goal was practical: group routers, gateways and switches so the service could remain reliable even when an individual element did not.
The work did not remain bound inside a thesis. Fan co-founded Rainfinity with his adviser, Shuki Bruck, and served as its chief technology officer. Rainfinity built software for network and file virtualization. Its early RainWall product applied the cluster idea to firewalls. By 2005, EMC had acquired the company.
An acquisition can end a founder story or widen it. Fan's widened. EMC put him in charge of a new software development center in China in 2006. The announced plan was to begin with about 100 developers and grow substantially. He was moving from the tight feedback loop of a research spinout into the coordination problems of a multinational engineering organization.
“Memory Machine is doing to persistent memory what VMware vSphere did to CPUs.”Charles Fan, describing MemVerge's abstraction layer
Chapter two
Building storage inside VMware
At VMware, Fan became a senior vice president and general manager. He founded the storage business unit that developed Virtual SAN, later styled vSAN, and led storage and big-data work. The assignment placed him at a useful boundary. VMware was known for turning physical compute into flexible pools. Storage still carried more of the old physical constraints.
Virtual SAN pulled local disks in servers into a software-defined shared system. The recurring move is hard to miss: take resources tied to individual machines, add a software layer, and present the application with something more resilient and easier to use.
Fan later served as chief technology officer of Cheetah Mobile, leading global technology teams. On paper, consumer mobile software looks like a departure from enterprise infrastructure. In practice, the role added another operating environment, one shaped by global scale and a different product tempo. When Fan returned to startup life, he brought the experience of both founding and running large groups.
One problem, four forms
Chapter three
The professor, the student and the postdoc
MemVerge's founding team made the career loop visible. Fan reunited with Bruck, his former Ph.D. adviser and Rainfinity co-founder. They were joined by Yue Li, a Caltech researcher in non-volatile memory who became MemVerge's chief technology officer. Bruck became chairman. Fan became CEO.
Fan has called himself a geek and an accidental entrepreneur. There is something useful in that description. The companies appear less like attempts to occupy a founder identity than commercial responses to technical transitions. Rainfinity emerged from fault-tolerance research. MemVerge began when a new kind of persistent memory looked ready to blur the old line between working memory and storage.
In 2017, Intel released an Optane solid-state drive built on 3D XPoint media. The MemVerge founders expected persistent-memory modules to follow. Their thesis was that new hardware would need a software layer before ordinary applications could benefit. MemVerge would combine fast DRAM, larger persistent memory and virtualization software into what it called Big Memory.
That choice also reunited three generations of the same technical conversation. Bruck had advised Fan's doctoral work on distributed reliability. Fan had taken virtualization from a Caltech spinout into EMC and VMware. Li brought current research on non-volatile memory systems. Their roles were not interchangeable: the professor chaired, the postdoc led technology, and the former student became the operator responsible for turning a laboratory opening into a company. MemVerge was Fan's second startup, but it did not require him to invent a new professional identity. It let him combine the research habits of his first company with the scale learned after its acquisition.
Choose a problem broad enough to survive a product cycle. Fan's products changed from firewalls to file systems, virtual storage, CXL pools and agent memory. The durable layer was the need to preserve state across changing infrastructure.
The company raised a $19 million Series B in 2020 led by Intel Capital, with strategic investors including Cisco Investments, NetApp and SK hynix. Disclosed funding across its rounds reached $43.5 million. The commercial pitch compared the company's Memory Machine layer to the way VMware had abstracted CPUs. Applications could see a larger, persistent pool without being rewritten around every hardware detail.
Hardware road maps rarely move in a straight line. The persistent-memory market did not unfold exactly as early advocates expected. MemVerge kept the word “memory” and expanded what it meant. The company worked on workload mobility in the cloud, transparent checkpointing, GPU utilization and CXL, an interconnect that lets processors and memory devices share resources beyond a single server.
That adaptation matters more than a tidy origin story. A company built around one component can become trapped by it. MemVerge treated the component as one route toward a larger mission: make active data abundant, available and manageable by software.
“Think Cloud and Think Memory. These will be the two most important forces in IT for the next decade.”Charles Fan, advice to young technology professionals in 2021
Chapter four
When the memory belongs to the agent
By 2025, the bottleneck had moved again. Large language models could generate fluent answers, but an agent often began each session without the user's working history. People had to repeat preferences, projects and prior decisions. MemVerge called this stateless amnesia and launched MemMachine, an open-source memory layer for AI applications.
The product stores and retrieves episodic, profile and procedural information across sessions, agents and models. That puts Fan's old abstraction pattern in a new place. The application should not have to care which machine holds a byte of persistent memory. The user should not have to care which model holds yesterday's conversation. A logical memory layer sits above the changing parts.
In April 2026, Fan appeared among the authors of a MemMachine research paper describing a ground-truth-preserving system for personalized agents. The paper reported results across long-conversation memory benchmarks and examined where retrieval accuracy improved. A quarter-century after his doctoral work, Fan was again attached to a document about preserving the right state across a distributed system. Only this time, the state included a person's context.
Watch: Supercharging AI infrastructure
Fan's January 2025 AI Field Day presentation connects agentic workloads, GPU orchestration and the infrastructure layer MemVerge was building.
The company's current language is more personal than its old data-center pitch. “The model serves you. Not the other way around,” Fan says on MemVerge's site. The phrase turns a systems architecture into a power relationship. If memory can travel across models, the user is less captive to any one of them. Continuity becomes a form of control.
It also changes the unit of persistence. In the early systems, the thing worth preserving belonged to an application: a connection, a file, a running job. In the agent model, memory is organized around a person and can be made available to several tools. That is a larger claim than faster infrastructure. It asks software vendors to treat identity and accumulated context as portable resources rather than exhaust from a single product session.
The operating layer
Persistence is also a team habit
Fan's technical work celebrates decoupling, but his companies also depend on unusually durable connections. Bruck was his professor, then his co-founder, then his co-founder again. Li's work at Caltech became part of MemVerge's technical base. The relationship graph survived job titles, an acquisition and two decades.
During pandemic remote work, Fan said MemVerge tried to protect that social layer with video happy hours. Colleagues had drinks, played games and talked about subjects other than work. His explanation was direct: a strong bond among team members is critical to shared success. For someone preoccupied with system continuity, the management instinct fits.
There is no guarantee that persistent AI memory becomes the final form of Fan's long-running problem. It probably will not. Technologies are temporary answers. What his career offers is a way to choose questions: find one important enough that each platform shift reveals another version of it.
Fan began by helping network services survive a failed box. He moved on to files, virtual machines, memory pools, cloud jobs and AI context. Each layer asks software to remember the thing people care about while the machinery underneath is replaced. The compelling part is not that he predicted every turn. It is that he kept recognizing the same constraint when it appeared in a new costume.