Context home windows of AI fashions, which point out the power of a mannequin to “bear in mind” info, have elevated over time. Nonetheless, researchers have recommended new methods to extend long-term reminiscence of AI fashions, as they typically can’t maintain context over a number of classes.
Nineteen-year-old founder Dhravya Shah is making an attempt to unravel issues on this space by constructing a reminiscence answer, known as Supermemory, for AI apps.
Initially from Mumbai in India, Shah began constructing consumer-facing bots and apps a number of years in the past. He even offered his bot that formatted tweets into handsome screenshots to the social media software Hypefury.
The founder, who was getting ready for an entrance examination to get into IIT (Indian Institute of Know-how), earned good cash from this sale and determined to maneuver to the U.S. to attend Arizona State College as a substitute.
After relocating, he challenged himself to construct one thing new every week for 40 weeks. Throughout a kind of weeks, he constructed Supermemory (which was initially known as Any Context) and put it on GitHub. At the moment, the software allowed you to talk along with your Twitter bookmarks.
The present model of the software extracts “reminiscences” or insights from unstructured information and helps the purposes perceive the context higher.
Shah secured an internship at Cloudflare in 2024, the place he labored on AI and infrastructure. He later labored as a developer relations lead on the firm. Throughout this time, advisors, together with Cloudflare CTO Dane Knecht, requested him to show Supermemory right into a product.
This yr, he determined to construct Supermemory full-time.
Supermemory, now described as a common reminiscence API for AI apps, builds a data graph based mostly on the information it processes, and personalizes the context for the customers. For example, it might assist querying throughout month-old entries for a writing or a journaling app, or trying to find an electronic mail app. As the answer permits for multimodal inputs, it may additionally enable a video editor to fetch related property from a library for a specific immediate.
The startup can ingest any sort of information, the corporate says, together with information, docs, chats, tasks, emails, PDFs, and app information streams. Its chatbot and notetaker function lets customers add reminiscences in textual content, add information or hyperlinks, and connect with apps like Google Drive, OneDrive, or Notion. There’s additionally a Chrome extension that permits you to simply add notes from a web site.
“Our core energy is to extract insights from any type of unstructured information and provides the apps extra context about customers. As we work throughout multimodal information, our answer is appropriate for all types of AI apps starting from electronic mail purchasers to video editors,” Shah stated.
Supermemory has secured seed funding of $2.6 million led by Susa Ventures, Browder Capital, and SF1.vc. The spherical additionally consists of particular person traders like Cloudflare’s Knecht, Google AI chief Jeff Dean, DeepMind product supervisor Logan Kilpatrick, Sentry founder David Cramer, and executives from OpenAI, Meta, and Google.
Shah stated at one level Y Combinator additionally approached him to affix one among its batches, however he already had traders on board, so the timing didn’t work out.
Joshua Browder, founder and CEO of “Robotic lawyer” startup DoNotPay, who runs Browder Capital as a solo GP, was impressed by Shah’s tenacity.
“I related with Dhravya over X, and what struck me was how rapidly he strikes and builds issues, and that prompted me to put money into him,” he stated.
The corporate has a number of current clients, together with a16z-backed desktop assistant Cluely, AI video editor Montra, AI search Scira, Composio’s multi-MCP software Rube, and actual property startup Rets. Plus, it’s working with a robotics firm to retain visible reminiscences captured by a robotic.
Whereas it is a slant towards shoppers, the app feels extra like a playground for builders to know extra concerning the software and doubtlessly use it of their workflows or their very own apps.
Supermemory has substantial opponents within the reminiscence area. Startups like Felicis Ventures-backed Letta and Mem0 (the place Shah labored for a short time) are constructing a reminiscence layer for brokers. Supermemory’s personal backer, Susa Ventures, has invested in Reminiscences.ai together with Samsung, which might faucet into 1000’s of hours of footage to get insights. Shah says that whereas these startups would possibly serve completely different industries and use instances, Supermemory will stand out as a result of it affords decrease latency.
“An increasing number of AI firms will want a reminiscence layer. Supermemory’s answer offers excessive efficiency whereas permitting you to floor related context rapidly,” Browder stated.
