Switch from ChatGPT to Claude mid-conversation and you'll notice something annoying: your AI just forgot everything. The coffee preferences you laboriously explained, the communication style you prefer, the projects you're juggling—all of it locked inside whichever app you trained first.
Egoist Machines, a startup emerging from Y Combinator, is betting users are tired of this memory fragmentation. Its answer: AI Passport, a user-controlled memory layer that sits between you and every AI assistant you touch.
The concept feels almost obvious in hindsight. What if your preferences, constraints, and personal context lived in one place you actually own, rather than scattered across ChatGPT's settings, Claude's conversation history, and Google's profile data? The San Francisco-based company opened a waitlist this past summer for what amounts to a portable identity system for the AI age—one that promises to work with everything from scheduling assistants to e-commerce chatbots.
General availability is expected later this summer, though the company hasn't locked down a specific date.
Permission Slips for AI Memory
At its core, AI Passport functions as a centralized profile. Any app or AI assistant can knock on the door and ask for information. Maybe your calendar tool needs to know your working hours. Maybe a shopping bot wants your size and budget constraints. Each request requires your explicit approval.
The architecture revolves around what Egoist calls "passes"—time-limited, scope-specific permissions. An app doesn't just get access to "your preferences." It asks for read access to your communication style preferences, for this specific task, for the next hour. You see the request. You approve or deny it. If approved, the app can read that single slice of information. If it wants to write a new memory—say, Claude learning you hate small talk—you review that update in an inbox before it gets saved.
According to developer documentation, the system supports over 800 connector sources. Whether all those connectors are live at launch remains unclear.
"Big tech is scary and eats your data for breakfast," co-founder David Khachaturov posted on LinkedIn when announcing the waitlist, framing the product as a privacy bulwark against the status quo of app-by-app memory silos.
The company's privacy policy—which took effect on July 22—explicitly states that stored memories won't be used to train AI models, a reassurance likely aimed at users wary of their personal details becoming training fodder. Special provisions address health and fitness data, and the policy details compliance with Google's API Limited Use requirements.
For developers, Egoist provides intake forms requesting use case descriptions, the specific data scopes needed, and whether access will be read-only or read-write. The company also references the Model Context Protocol (MCP), an emerging standard for AI context sharing, suggesting compatibility with that framework.
The Founders
Egoist Machines is a two-person show. CEO Erin McGurk and co-founder David Khachaturov both graduated from Cambridge before relocating to San Francisco for Y Combinator's most recent summer batch. McGurk built a substantial following—over a million—as @erinmerylstudy before pivoting to the startup world. She and Khachaturov reportedly met at Cambridge during a Caledonian dancing event, according to a BusinessCloud profile published in mid-July.
Khachaturov's resume leans academic. He earned a PhD in Computer Science from Cambridge, where he became the youngest-ever Fellow at St Catharine's College at 25. Before Egoist, he worked as a machine learning researcher at OpenAI (contributing to PaperBench) and later as lead AI security researcher at FrontierOne.
The company also maintains open-source projects. LodeDB is an embedded vector database pitched as a "durable memory backend," with benchmarks against LangChain and LlamaIndex defaults. Another project, etchplan, compiles agent workflows into deterministic, auditable routines—tools that hint at the technical foundations underpinning AI Passport.
A Crowded, Chaotic Field
Egoist isn't alone in pursuing AI memory infrastructure. Mem0 offers persistent memory for agents and apps, targeting developers with LangChain and CrewAI integrations. MemoryLake markets a "Memory Passport" aimed at cross-platform portability. Tools like Debo, WeaveIt, and EidosOmni each approach personal AI memory from slightly different angles. Priver AI even offers a hardware-rooted solution with on-device vaults.
The "passport" metaphor itself seems to be having a moment. Research papers from earlier this year explored agent identity and authorization frameworks (Open Agent Passport), and DigiCert published a Q&A arguing that "AI agents need passports, not passwords." Most of these efforts, though, focus on agent identity rather than user-owned memory.
Academic research is catching up as well. A paper published in June titled "Beyond Similarity: Trustworthy Memory Search for Personal AI Agents" flagged risks in memory systems, including cross-domain leakage and jailbreaks, while evaluating tools like Mem0 and MemOS. Another July paper proposed an end-to-end memory stack for personal AI agents.
What Egoist is pitching businesses, however, feels more immediately practical: e-commerce personalization without forcing customers through endless quizzes (the Passport already knows sizes, budgets, style constraints), or support teams maintaining unified customer memory across tickets so new cases start informed rather than ignorant.
Whether those use cases resonate beyond pitch decks remains an open question.
The Control-Convenience Tension
Pricing for end users is simple: free, forever, according to company posts from mid-July. Developer and enterprise pricing hasn't been disclosed. Neither have detailed storage architecture specs beyond what's outlined in privacy documents and terms of use.
Public API documentation remains limited to a developer intake form, with connectors intended as per marketing materials. No seed, angel, or venture round amounts or cap table details have been publicly announced. The company hasn't disclosed user counts or waitlist size.
For now, interested users can sign up at ego.ist and wait.
The bigger question isn't technical—it's behavioral. Does anyone actually want this much control over their AI memory? The inbox approval model offers granular privacy, certainly. But it also introduces friction. Every time an app wants to learn something new about you, you have to review and approve. That might feel empowering at first. It might also feel exhausting.
AI Passport's fate likely hinges on that tension. If the founders can thread the needle—making control feel liberating rather than burdensome—they might build the connective tissue our fractured AI ecosystem badly needs. If not, they'll just be another layer of complexity in an already complicated relationship between humans and their digital assistants.
The waitlist will tell part of the story. Developer adoption will tell the rest.
