Category: AI
Headline & intro
Smart glasses that can’t remember what they saw are little more than fancy cameras. The same goes for home robots that forget your living room the moment they roll into the kitchen. The next wave of AI hardware will live in the physical world, and that world is messy, continuous and visual. That’s exactly the gap Memories.ai is trying to occupy with its “visual memory layer” for wearables and robots. In this piece, we’ll unpack what they’re building with Nvidia, why it matters strategically, and whether a startup can realistically own this layer before Apple, Meta or Google do.
The news in brief
According to TechCrunch, U.S. startup Memories.ai has announced a collaboration with Nvidia, revealed at Nvidia’s GTC 2026 conference, to build infrastructure that lets AI systems store and recall visual memories.
Memories.ai, founded in 2024 by Shawn Shen and Ben Zhou after working on Meta’s Ray-Ban smart glasses, uses Nvidia’s Cosmos-Reason 2 vision-language model and Nvidia Metropolis video search/summarisation tools as core building blocks. The company has developed a large visual memory model (LVMM) that embeds and indexes video so it can later be searched and reasoned over.
Memories.ai has raised around $16 million to date, including an $8 million seed round plus an $8 million extension led by Susa Ventures with investors such as Seedcamp and Crane Venture Partners. The startup has also signed a partnership with Qualcomm so its second-generation LVMM can run on Qualcomm processors, and is already working with unnamed major wearable makers.
Why this matters
Wearables and robots currently suffer from a kind of “goldfish AI”: powerful perception in the moment, almost no durable memory of what happened five minutes ago, let alone last week. Memories.ai is betting that whoever solves that gap becomes the default middleware for every device that needs persistent understanding of the physical world.
In practice, this “visual memory layer” is a specialised form of vector database plus reasoning engine tuned for video and images. If it works, a factory robot could ask, “Have I seen this part misaligned before?”; AR glasses could answer, “Where did I leave my keys yesterday?”; a field technician’s headset could surface the exact wiring diagram they looked at on the last visit to this substation.
The immediate winners are:
- Nvidia, which gets another showcase for its vision models and Metropolis stack and further locks AI startups into its ecosystem.
- Qualcomm and other edge chip vendors, who gain a concrete use case for on-device AI that’s not just running a generic LLM.
- Big wearable OEMs, who can outsource a messy, infra-heavy problem instead of building their own memory stacks from scratch.
The potential losers are:
- General-purpose LLM platforms that stay text-centric and miss the embodied AI shift.
- Smaller hardware startups that can’t match this level of infra and may be forced onto whichever memory layer the dominant platforms choose.
Most importantly, visual memory turns continuous video streams from a liability (storage, bandwidth, privacy risk) into something queryable and economically valuable. That flips the entire cost-benefit equation for putting cameras everywhere.
The bigger picture
Memories.ai sits at the intersection of three powerful trends.
First, AI models have quietly been gaining long-term memory. As TechCrunch notes, OpenAI, Google Gemini and xAI have all rolled out text-based memory features since 2024. But those systems primarily remember what you typed, not what you and your devices saw. Text memory is relatively easy: it’s structured, compact and aligns nicely with existing databases. Visual memory is messier, higher bandwidth and more ambiguous—but it’s exactly what embodied AI needs.
Second, we’re in a new wave of AI wearables and agent devices. From Meta’s Ray-Ban glasses to experiments like Humane AI Pin and Rabbit’s devices, the hardware is converging on the idea of a persistent assistant that “lives with you”. All of them hit the same wall: without robust, queryable memory, the wow-factor demos evaporate quickly. Memories.ai is effectively proposing to be the “memory OS” these products are missing.
Third, robotics and automation are moving from static, pre-programmed systems to adaptive agents. Industrial and logistics robots increasingly need to remember environments, anomalies and rare events over long timeframes. Historically, each robotics company hacked together its own perception+logging stack. A standardised visual memory layer could do for robot perception what cloud databases did for web apps.
There’s historical precedent too. Early “lifelogging” experiments—think Google Glass or niche cameras like Narrative Clip—failed in part because the captured media wasn’t searchable or useful. The promise now is that we finally have the models and silicon to turn that firehose into something structured enough to query, while running much closer to the edge.
The strategic question is whether this becomes a horizontal infrastructure market (where a company like Memories.ai can thrive), or whether platform giants fold visual memory into their own stacks and commoditise everything else.
The European / regional angle
For Europe, a visual memory layer is both an industrial opportunity and a regulatory minefield.
On the upside, this technology aligns perfectly with European strengths in robotics and manufacturing. Think of German automotive plants, Swiss logistics hubs, or Nordic warehouse automation. A standardised way for robots and inspection systems to remember what they’ve seen could drive productivity without replacing entire fleets. European OEMs are historically comfortable with buying specialised middleware if it’s robust and well-governed.
On the downside, Europe has the strictest privacy rules in the world. Continuous video capture that is then embedded, indexed and queried squarely touches GDPR (biometric data, bystander rights, purpose limitation) and the EU AI Act, which explicitly targets biometric surveillance and high-risk uses. A generic “record everything” wearable with rich visual memory would be extremely hard to deploy in public spaces without aggressive on-device processing, minimisation and consent mechanisms.
This is where European players could differentiate: by building privacy-preserving visual memory. That could mean architecting systems where embeddings never leave the device, or where the memory layer is tightly bound to enterprise contexts (factories, warehouses, hospitals) with clear governance, rather than consumer lifelogging.
For European startups in vision AI, there’s a choice: integrate with U.S.-centric stacks like Memories.ai + Nvidia, or push for EU-grown alternatives (leveraging initiatives around sovereign cloud and open models). Either way, Brussels will shape what “acceptable” visual memory looks like long before the market matures.
Looking ahead
Over the next 12–24 months, expect Memories.ai to behave less like a consumer brand and more like deep infrastructure: SDKs, APIs and reference designs for OEMs. If they succeed, end users may never see the logo, only features like “Recall what my glasses saw last week.”
Key things to watch:
- Integration depth – Do major wearable makers and robot OEMs expose this as a flagship capability, or keep it as a quiet internal tool?
- On-device vs. cloud – How much of the LVMM stack can realistically run on Qualcomm-class chips, and what must still happen in the cloud for heavier reasoning?
- Regulatory test cases – The first big privacy controversy involving visual memory (for example, glasses recalling faces in public) will set norms and perhaps case law.
- Platform response – If Apple, Meta, Google or OpenAI announce their own visual memory layers tightly coupled to their ecosystems, the window for an independent layer narrows fast.
The bigger technical risk is data quality. The LUCI data-collection wearables described by TechCrunch give Memories.ai a bespoke dataset, but ensuring diversity across environments, cultures and edge cases is hard and expensive. If the model underperforms outside its training distribution, OEMs will revert to in-house solutions.
Still, the direction of travel is clear: the AI that matters in five years will be the AI that remembers and reasons over what it has seen, not just what we have typed.
The bottom line
Visual memory is shaping up to be the missing infrastructure layer for AI in the physical world. Memories.ai’s Nvidia-backed push puts it early in a race that will decide who owns that layer: independent infra players or the platform giants. The technology will collide head-on with European privacy rules, but it also unlocks enormous value in industry, robotics and serious wearables. The real question for readers is simple: are you ready for devices that never forget what they’ve seen—and who should be allowed to own those memories?
