There is a certain kind of technology company whose whole existence is a bet on a single, slightly contrarian idea. MemVerge's idea is that memory - the fast, expensive, forgetful stuff that sits between the processor and the disk - is more important than everyone treats it, and that if you virtualize it properly, interesting things happen. This is not an obviously thrilling premise. It is, however, an idea that keeps turning out to be right at inconvenient times for everyone who bet against it.
The company was founded in 2017 in Milpitas by three people with unusually intertwined résumés. Charles Fan, the CEO, had earlier co-founded a startup called Rainfinity with Shuki Bruck, a Caltech professor who happened to be his PhD advisor; EMC bought it. Fan then went to VMware and started the storage business unit that built VSAN into a billion-dollar product. Bruck became MemVerge's chairman. Yue Li, the CTO, had been Bruck's post-doc. So the founding team is, essentially, an advisor, his student, and his student's student, which is either a charming academic lineage or a slightly recursive org chart, depending on your temperament.
The original product was called Memory Machine, and its trick was to virtualize different types of memory - ordinary DRAM and Intel's then-new Optane persistent memory - into a single pool that software could tier, replicate, snapshot and recover at memory speeds. MemVerge gave this a name, "Big Memory Computing," because naming a category is half the work of creating one.
When the hardware died
Here is the inconvenient part. MemVerge's early bet leaned heavily on Optane. Intel discontinued Optane. In the normal course of things, this is the sort of event that quietly ends a company - you have built a business on a chip, and the chip is now a museum piece.
MemVerge's response was to notice that its actual expertise was never really about Optane. It was about memory: tiering it, pooling it, moving it around without applications noticing. And it turned out there was a much larger customer arriving who cared about exactly that. That customer was generative AI.
"AI Memory: The Next Frontier."
— MemVerge company taglineAI has two memory problems, and MemVerge decided to attack both. The first is hardware. GPUs are the most expensive idle asset in a modern data center; they frequently sit waiting on memory rather than computing, which is a bit like buying a fleet of Ferraris and parking them. Memory Machine's tiering, pooling and transparent checkpointing aim to keep those GPUs fed. A joint demo with Micron reported 77% higher GPU utilization and more than doubled the speed of OPT-66B batch inference. A financial-services customer reportedly doubled GPU utilization outright.
The second memory problem
The second problem is stranger and more human. Large language models have no memory. Every conversation begins from zero; the model that felt like it knew you yesterday has, today, never heard of you. MemVerge calls this "stateless amnesia," and in September 2025 it launched a product aimed squarely at curing it: MemMachine, an open-source AI memory layer.
MemMachine gives agents three kinds of memory, borrowed from how psychologists describe human memory: episodic (what happened), profile (who you are), and procedural (how to do things). It works across the major models - OpenAI, Claude, Gemini, Grok, Llama, DeepSeek, Qwen - and on any cloud or on-prem. The company reported a LoCoMo recall benchmark of 84.87% and claims roughly 50% lower token costs, on the logic that remembering something is cheaper than re-explaining it in every prompt.
MemMachine retains episodic, personal and procedural knowledge across sessions, models, agents and environments - ending the stateless amnesia of today's LLMs.
— MemVerge, on the MemMachine launchThe open-core wager
The MemMachine strategy is a familiar one in infrastructure: give the core away and sell the enterprise. The open-source project lives at memmachine.ai, with a free playground; the commercial side at memverge.ai adds scalability, security, compliance, orchestration, observability and a "White Glove" service tier reported around $2,500 a month. The open project is the front door. The business is what happens after developers walk through it.
Meanwhile the hardware thread continues. At SC25 in late 2025, MemVerge and XConn Technologies demonstrated a CXL - Compute Express Link - memory pool that offloads and shares the KV cache across GPUs and CPUs, claiming greater than 5x performance over SSD-based caching while cutting cost of ownership. This is the same company idea it started with, wearing new hardware: pool the memory, share it, and stop letting expensive silicon wait.
MemVerge is not a large company - roughly 66 people, about $43.5 million raised across a 2020 Series A backed by Intel Capital, Cisco, NetApp, SK hynix, Lightspeed and others. It competes in a crowded new category of AI-memory startups such as Mem0 and Zep, and against the vector-database stacks everyone already uses. Whether it wins is genuinely unknown. But there is something worth noticing in a company that has now been correct about the same unglamorous idea across three hardware generations, and simply kept waiting for the rest of the industry to arrive.