Samsung backs a video AI startup that can analyze thousands of hours of footage

Many AI tools can look at a video today and summarize what is going on, but things become a bit tricky when you ask models questions about multiple videos and footage spanning many hours.

This is a big limitation for security companies that want to use AI to scrub through thousands of hours of footage from different cameras, as well as marketing companies that want to study different video campaigns and product shoots.

Memories.ai wants to tackle that problem with its AI platform that can process up to 10 million hours of video. For companies with a lot of video to analyze, the startup wants to provide a contextual layer, complete with searchable indexing, tagging, segments and aggregation.

Its co-founder Dr. Shawn Shen was a research scientist at Meta’s Reality Labs while he was pursuing his PhD, and his counterpart Enmin (Ben) Zhou worked at Meta as a machine learning engineer.

“All top AI companies, such as Google, OpenAI and Meta, are focused on producing end-to-end models. Those capabilities are good, but these models often have limitations around understanding video context beyond one or two hours,” Shen told TechCrunch.

“But when humans use visual memory, we sift through a large context of data. We were inspired by this and wanted to build a solution to understand video across many hours better,” he said.

Towards that goal, the company has now raised $8 million in a seed funding round led by Susa Ventures, and with participation from Samsung Next, Fusion Fund, Crane Ventures, Seedcamp, and Creator Ventures. Shen said the company initially aimed to raise $4 million, but ended up with an oversubscribed round because of investor interest.

“Shen is a highly technical founder, and he is obsessed with pushing boundaries of video understanding and intelligence,” said Misha Gordon-Rowe, a partner at Susa Ventures. “Memories.ai can unlock a lot of first-party visual intelligence data with its solution. We felt that there was a gap in the market for long context visual intelligence, which attracted us to invest in the company,” he added.

Samsung Next had a slightly different thesis: the investment arm of Samsung sees Memories.ai’s solution being useful for consumers.

“One thing we liked about Memories.ai is that it could do a lot of on-device computing. That means you don’t necessarily need to store video data in the cloud. This can unlock better security applications for people who are apprehensive of putting security cameras in their house because of privacy concerns,” said Sam Campbell, a partner at Samsung Next.

Memories.ai says it uses its own tech stack and models to perform analyses. First, it removes noise from videos and passes the output through a compression layer to only store important data. Then there is an indexing layer, which makes the video data searchable (using natural-language queries) with segmentation and tags. There is also an aggregation layer that summarizes data from the index, helping create reports.

Currently, the startup caters to two kinds of companies: marketing and security. Marketing companies can use the startup’s tools to look up trends related to their brands on social media, and identify what kind of video they want to make. Memories.ai also provides tools for marketers to create those videos.

The company is also working with security companies to help them analyze security footage to determine potentially dangerous actions by people in the videos by reasoning through patterns.

Currently, companies working with Memories.ai need to upload their video library to the platform to have it analyze clips. But Shen said that in the future, his clients will be able to create a shared drive and sync content more easily. The plan is to enable customers to ask questions like: “Tell me all about people I interviewed in the last week.”

Shen envisions an AI assistant that can gain context on a user’s life via their photos or when they activate smart glasses. He also sees the technology playing a role in training humanoid robots to do complex tasks or helping self-driving cars remember different routes.

The company currently employs 15 people, and it plans to use the fund to augment its team and improve its search.

Memories.ai is going up against similar startups, like mem0 and Letta, which are working on providing a memory layer for AI models, though they offer limited video support at the moment. It also has to contend with companies like TwelveLabs and Google, which have been working on helping AI models understand videos.

Shen, however, feels his company’s solution is more horizontal, which would let it work with different video models as well.