‘Large Memory Models’ pitch emerges
A startup called Engramme introduced the idea of 'Large Memory Models' as an alternative to classical RAG, pitching a memory‑centric architecture for perfect recall of emails and conversations. The founders, with an academic background, argue this approach mimics human memory to serve long‑horizon retrieval tasks without the same vector‑search choreography. It’s an explicit product attempt to reframe how persistent knowledge and recall are implemented in enterprise workflows. (x.com)
Most company chatbots still work like a librarian with bad eyesight: they chop your files into snippets, turn those snippets into numbers, and hope a vector search pulls back the right pieces when you ask a question. Microsoft, Amazon Web Services, and Google all describe retrieval-augmented generation that way in their official docs. (learn.microsoft.com) (docs.aws.amazon.com) (cloud.google.com) That works well for manuals, policies, and product pages, but it gets awkward when the thing you want is buried across months of email threads, calendar invites, and half-finished conversations. Azure’s own guidance says classic retrieval-augmented generation depends on indexes, embeddings, and chunk retrieval, which is a lot of machinery for “what did Sarah promise in the meeting before the offsite?” (learn.microsoft.com 1) (learn.microsoft.com 2) A startup called Engramme is pitching a different answer: don’t treat your digital life like a pile of documents to search, treat it like a memory to recall. Its website says the goal is “searchless, promptless recall” for every person met, conversation had, and place visited. (engramme.com) The company is not coming out of a normal software background. Engramme’s about page says co-founder and chief executive Gabriel Kreiman spent almost 20 years as a Harvard professor researching artificial intelligence and memory, and co-founder and chief technology officer Spandan Madan earned a doctorate at Harvard University and the Massachusetts Institute of Technology Computer Science and Artificial Intelligence Laboratory. (engramme.com) That academic angle shows up in the product pitch. Engramme says it is building “novel AI models of human memory,” and it links that claim to research on how neuronal circuits encode memories and to work predicting what people remember from movies. (engramme.com) The phrase the founders are using is “Large Memory Models.” Bloomberg reported on April 10, 2026 that Kreiman described the startup on LinkedIn as building large memory models that can access data across a person’s digital life and surface relevant information automatically, without the user prompting for it. (bloomberg.com) Engramme is also trying to prove that “memory” is a distinct problem, not just another search feature. On its site, the company says one study collected 1,940 personal memory questions from 134 participants during daily life to map what people most often fail to recall. (engramme.com) That is the real bet here: enterprise knowledge tools have spent three years teaching models to fetch the right document, while Engramme wants a system that behaves more like a colleague who remembers the thread, the people, and the missing detail without being told where to look. Its website calls that “omniscient AI to augment human cognition,” which is a much bigger ambition than a better chat-with-your-PDF tool. (engramme.com) Investors are at least listening. Bloomberg reported that Engramme launched out of stealth in March 2026 and is in talks to raise about $100 million at a valuation discussed as high as $1 billion, though the report said the talks were early and terms could change. (bloomberg.com) Whether “Large Memory Models” becomes a real category or just a sharp piece of branding depends on one hard test: can it recall the right fact from your messy life more reliably than the retrieval-augmented generation stacks already shipping from cloud vendors. Right now, the novelty is not that Engramme wants better answers, but that it is trying to replace search itself with something closer to machine memory. (learn.microsoft.com) (engramme.com)
