First Person Labs came out of stealth on 17 November with roughly $10 million in seed funding led by the infrastructure team at Andreessen Horowitz, and a product that takes a contrarian position on how AI should remember people. Instead of each assistant keeping its own notes, Second builds a single continuously updated model of its user, stored where that user can inspect it, and serves it to whichever agents they authorise.
The company was founded about seven months ago by Ken Liu and Erik Chi, both PhD students in the Stanford Artificial Intelligence Laboratory. Liu researched AI and data privacy at Stanford and Google DeepMind, and Chi worked on network security and privacy at Michigan. Mozilla Ventures participated in the round, along with individual backers including Jeff Dean and Percy Liang.
The mechanics are deliberately unglamorous. Second runs in the background on a Mac and learns from computer activity at the operating-system level, without the user filing anything. It keeps that state in a folder that can be inspected, edited, backed up or moved, streams it to a private cloud, and then exposes the resulting context over local and cloud APIs or MCP. Users bring their own Claude or ChatGPT subscription, their own API key, local inference, or buy private inference from the company.
Access controls are per app and per agent, so a user decides which parts of their context a given tool may read. First Person Labs says it does not see, train on, log or sell user data unless a customer opts into hosted inference or Second Cloud, and that it offers zero data retention and confidential computing. It also intends to open-source key components so that people can self-host.
"Any interface you talk to knows you," Liu said, describing the product as being like a biographer, someone who follows you around and sees everything. The a16z investment post makes the strategic case bluntly: whoever controls that model will have considerable influence over what your agents do for you. That is a direct challenge to the memory features shipping inside ChatGPT, Claude Code and Meta Muse, each of which accumulates a private, non-portable picture of the same person.
The gaps are real. It is not disclosed how much computer activity Second records, how that store is protected against an attacker who reaches the machine, or what the product costs after early access. Background agents that read everything on a laptop are also a large prompt-injection surface, and operating-system permissioning limits what any independent developer can see. On macOS in particular, the platform owner controls the signals a product like this depends on.
The interesting claim is architectural rather than technical. If memory becomes a layer that follows the user instead of a feature that belongs to each vendor, then switching assistants gets cheap and the lock-in that memory creates weakens. That is a good outcome for users and a bad one for anyone whose retention strategy is a growing private profile. First Person Labs is betting that portability wins, and that a user-owned context store is more trustworthy than six vendor silos. The bet is credible; the security engineering that would justify it is not public yet.
Comments (0)
Log in to join the discussion
Log InNo comments yet