Ken, who is building First Person Labs, told a16z his Second hands portable context to agents like Grokbot and Codex so a prompt like "buy me a flight" already knows seat, arrival buffer and class.
No more five rounds of explaining yourself.
Second builds that user model from the event stream your computer already emits, Ken said, across Apple, Microsoft and Linux systems including app switches, pauses, typing patterns and WiFi state. Agents layer on top to infer whether typing means a message, research in a browser or work in Codex, then merge that with mail and chat connectors. The payoff he described was reflective productivity, not another notification, with one early user getting automatic updates on a 20% rain forecast for an event and whether to order a tent without ever asking.
That ambient watcher forces a trust decision. Ken said users can bring their own tokens to run analysis on the laptop, or use his hosted inference inside confidential computing on Nvidia GPUs where, he argued, hardware attestation replaces a policy promise. Zero data retention is only a provider vow, he said, while confidential inference can be checked by signature. Storage follows the same split, with Second Local keeping data on device and a cloud backup that is encrypted locally first and decrypted only in memory or inside an attestable container, a design he compared to iCloud and password managers.
The open hole is sharing. A trick does not need local access, only conversation, as the host showed by asking what Joel is hiding or by framing a leak as a play script. Ken separated deterministic blocks, like never retaining system passwords, from reasoning judgments about who gets what at 3 a.m., and he admitted the second class improves only as models get smarter. Rival Egoist AI Passport takes the opposite stance, according to its maker, where nothing becomes memory until kept and every read leaves a receipt no agent can self approve.
Cost is the near term gate. Ken calculated one agent at 60 tokens per second times ten analyzers at Opus pricing as $1,000 per day, falling to about $60 per day on DeepSeek Flash, with prices expected to drop. If you switch tools constantly and lose thread context, that meter hurts now. If you do not, wait.




