Egoist Machines, a two-person startup backed by Y Combinator, has built what amounts to an identity layer for artificial intelligence. Their product, AI Passport, lets users decide which fragments of their digital lives ChatGPT or Claude gets to see, and when those assistants have to forget.
The concept resembles OAuth, the login protocol that lets you sign into apps using Google or Facebook credentials, except the currency here is personal context. An AI tool requests permission to access specific categories of user data, explains why it needs that information and for how long, then waits for approval. Users can revoke access whenever they choose, according to company documentation.
At the center sits an account on ego.ist, where information streams in from connected services and waits in an inbox as proposals. A Whoop fitness tracker logs a workout. Google Calendar reveals a pattern of early-morning meetings. A preference mentioned to ChatGPT gets captured. None of it becomes a stored memory until the user says so. When apps want to read those memories later, they have to ask for exact categories by name.
Receipts track what each application actually used. Permissions can be yanked mid-session. "Think of your EGOIST AI Passport like camera permissions," wrote founder Erin McGurk on LinkedIn when the product launched.
The startup claims integrations with ChatGPT, Claude, Gemini, Perplexity, Spotify, Google Calendar, Gmail, Strava, Notion, Slack, and access to more than 800 connector sources through its developer platform, though it hasn't published a full list. A ChatGPT plugin appeared in OpenAI's directory as version 1.0.0.
Who's Building It
McGurk studied Land Economy at Cambridge and was involved with Egoist Machines during their Y Combinator acceptance. Her co-founder, David Khachaturov, earned a PhD in computer science from Cambridge, where he held prestigious academic positions, including at St Catharine's College, according to the company's Y Combinator profile. Before Egoist Machines, he worked as a machine learning researcher at OpenAI and led AI security research at FrontierOne.
The company remains just the two of them, with Harshita Arora listed as their primary Y Combinator partner. Perhaps that lean structure is intentional for a startup handling sensitive personal data; fewer people means fewer potential points of failure.
Under the Hood
AI Passport runs on LodeDB, an open-source embedded vector database that Egoist Machines describes as exceptionally fast and compact. Each user gets an isolated data store protected by cryptographic access controls.
When a question comes in, the system routes it to what the company calls "an open model running on GPUs we control, inside a container that cannot reach the internet." Sessions get sealed with AES-256-GCM encryption. A feature called Memory Lock can further protect stores using keys derived from user passkeys. Revoking category-level access cuts off data immediately at the next request. Deleting data doesn't just remove files; the system shreds the encryption keys.
The architecture reflects a clear thesis: users need technical guarantees, not just policy promises.
The Privacy Pitch
Egoist Machines pledges never to sell memories, share them with third parties for their own purposes, or use them to train models. The company says its access to Google data is read-only and on-demand, compliant with Google API Services User Data Policy. Health and fitness information requires explicit consent and stays out of training pipelines.
"Until now, there has not been a permission-based system for sharing personal info with AI assistants," McGurk said when announcing the product. The implication: people deserve control, even if they've mostly accepted its absence.
A Crowded, Shifting Landscape
ChatGPT and Claude both expanded their memory capabilities over the past year. OpenAI added a "dreaming" feature to curate what assistants remember; Anthropic released tools to import and export memories. But those systems trap context inside a single assistant. Personal.ai offers a user-owned memory stack with API access. Mem builds a unified workspace across notes and meetings. Neither positions itself as authentication infrastructure spanning multiple AI services.
The personal AI memory sector has seen turbulence. Meta acquired Limitless AI, and Rewind shut down its Mac app, developments that suggested the independent middleware play might be harder than it looked. Academic papers on portable agent memory protocols multiplied, and the EU's Digital Markets Act began requiring large platforms to offer continuous, real-time data portability to authorized third parties. Interoperability went from a wish to a regulatory requirement.
The Data Transfer Initiative, backed by Apple, Google, Meta, and Microsoft, recently launched an AI Conversation Schema project for exporting chat histories across services. That creates a standards foundation for products like AI Passport, though whether the big platforms will genuinely open their gardens remains an open question.
Getting In
Egoist Machines initially ran a waitlist before opening general access to let users test the web version ahead of mobile apps. The product site now allows anyone to claim a handle and log in, though the company hasn't disclosed whether it has paying customers or how it plans to make money beyond Y Combinator's standard backing.
The startup says it has "distribution and deep-integration partners collectively reaching millions," but hasn't named them publicly. That could mean serious traction or careful phrasing around early conversations, the sort of ambiguity common in venture-backed companies still finding product-market fit.
"Data is the new oil," Khachaturov told BusinessCloud, deploying a metaphor that has worn thin from overuse. Then he added a twist: "Except this time every individual is swimming in it."
The question is whether individuals will pay attention long enough to care who controls the pools.
