Engramme’s large‑memory beta
Engramme announced a public beta for a Large Memory Models API that claims a non‑Transformer architecture for persistent, human‑like memory across apps like Gmail, Zoom and Slack—promising zero hallucinations and no external search or prompting. The announcement frames persistent memory as a distinct architectural primitive for assistants rather than a prompt‑engineering trick (x.com).
Most artificial intelligence assistants still work like a short-term note on a desk: they can use what is in the prompt, and then they forget it when the window closes. Anthropic’s documentation calls that working space a “context window,” and warns that accuracy can degrade as more text is stuffed into it. (anthropic.com) That is why companies bolt on retrieval systems, which are basically search engines for your own files and chats. The assistant does not truly remember your last Zoom call or an old Gmail thread unless another system finds it and pastes it back into the prompt. (anthropic.com) Engramme is pitching a different idea: memory as its own layer, not a bigger prompt. On April 9, 2026, the company opened a public beta for a Memory Application Programming Interface that developers can embed inside apps and services. (testingcatalog.com) The company says this layer can surface a user’s relevant past interactions without a search box and without the user asking first. TestingCatalog reports Engramme is targeting products tied to Gmail, Zoom, WhatsApp, Slack, Google Docs, and wearables including Meta glasses. (testingcatalog.com) The boldest claim is architectural, not cosmetic. Engramme says its “Large Memory Models” do not use the transformer design that underpins most modern language systems, and instead retrieve memories linked to real events, conversations, documents, and interactions. (testingcatalog.com) That is where the “zero hallucinations” line comes from. Engramme’s argument is that a memory layer built to fetch stored events should not invent a meeting, email, or document that never existed, because it is retrieving records rather than generating the most likely next words. (testingcatalog.com) The company’s own website frames this as “searchless, promptless recall” and says its long-term goal is to help people remember every person, conversation, and place they have encountered. The same site says users can sign up for the beta now. (engramme.com) Engramme is also trying to root that pitch in neuroscience rather than only software marketing. Its research page says the team collected 1,940 personal memory questions from 134 participants during daily life, and says related work has been accepted for publication in Nature Human Behaviour. (engramme.com) The founders match that story. Engramme’s about page lists Gabriel Kreiman, a Harvard Medical School professor who has studied memory and artificial intelligence for almost 20 years, as co-founder and chief executive, and Spandan Madan, who earned a Doctor of Philosophy at Harvard University and Massachusetts Institute of Technology Computer Science and Artificial Intelligence Laboratory, as co-founder and chief technology officer. (engramme.com) There is still a big gap between “memory layer” and “universal memory.” The public reporting so far does not show benchmark data, error rates, or independent tests for the claim that hallucinations are structurally impossible, so for now the beta is a product launch with an unusually ambitious theory attached to it. (testingcatalog.com)
