01 Trigger Event

a16z announced an investment in First Person Labs, whose product, Second, aims to continuously maintain a personal model on users’ devices and provide that context to the agents they choose.

This is not just another chatbot that “knows you better.” More precisely, it is an infrastructure bet: for personal agents to work across applications, someone must accumulate, refresh, authorize, and distribute the accumulated representation of who the user is.

a16z’s core argument is that Second’s background engine runs on the device by default, while the personal model remains inspectable and portable. Users can also choose which intelligence drives it and which information other applications and agents may access.

The founding team has relevant credentials: Ken Liu has studied AI and data privacy at Stanford and Google DeepMind, while Erik Chi studied network security, privacy, and censorship resistance at Michigan. Both previously participated in the Open Anonymity Project.

I have not used Second internally, so my assessment of its product capabilities and technical boundaries is based primarily on the investment thesis.

02 The Real Meaning

The issue is not whether AI can remember users, but who owns that memory.

Today, context is mostly locked inside ChatGPT, Claude, Cursor, or a session within some other SaaS product. Each product tries to build a user profile from its own interactions, but what users receive is not a portable personal context; it is a collection of disconnected data silos.

The more agents users adopt, the worse this fragmentation becomes. A user explains coding standards in Claude Code, then has to repeat them in Cursor. The CRM knows the customer’s preferences, the calendar knows their time constraints, and the browser agent can see neither. So-called autonomous agents often end up as isolated automation scripts that operate independently.

Second is betting on the context aggregation layer: continuously maintain a personal model, then inject authorized context into different agents. If this position holds, it could control a critical distribution layer in the agent era—not by selling tokens or completing tasks for users, but by determining what agents can see while executing those tasks.

Its moat could come from three areas.

First, long-term accumulation. Only by continuously absorbing users’ behavior, relationships, habits, and intentions can a model evolve from a static profile into living personal state.

Second, permission architecture. Who can read the context, how much they can read, when they can read it, and whether access can be revoked are harder to replicate than model intelligence and are more likely to create switching costs.

Third, trust. If an agent makes a costly decision because of incorrect context, users will rapidly tighten permissions. Personal context is not an ordinary preference; it is a high-leverage control system.

However, a16z did not disclose ARR, retention, active-user numbers, or the accuracy, latency, and memory-update mechanisms of the on-device model. The investment article presents a thesis, not evidence of traction. It would be premature to conclude that Second has achieved product-market fit.

03 Historical Analogy / Structural Comparison

This resembles the mobile identity aggregation enabled by the iPhone in 2007 more than an ordinary product launch.

The iPhone’s significance was not its touchscreen browser, but its consolidation of phones, email, maps, cameras, calendars, and third-party apps within a unified OS and account system. For the first time, users had a relatively continuous digital identity, while third-party developers built an ecosystem around iTunes distribution.

Agents introduce a similar structural shift: AI is moving from an app that users actively open to a continuously operating software actor. Once an agent can send messages, modify code, schedule meetings, or handle finances, its context can no longer come from a single interaction. It must come from personal state maintained across devices and applications.

Historically, browsers consolidated navigation at Google, but search and advertising ultimately gave Google control of distribution. A comparable layering could emerge in the agent era:

  • Model providers supply intelligence.
  • Tool platforms provide execution surfaces.
  • Context layers maintain personal state.
  • Applications deliver workflows and outcomes.

What ultimately gets valued may not be who owns the most agents, but who becomes the default interface between personal state and agent execution.

There is also a counterintuitive point: the context layer need not belong to a model company. Model capabilities will gradually become commoditized, whereas user authorization, data structures, and cross-application integration sit closer to the intersection of identity and operating systems. Model companies can certainly integrate upward, but First Person Labs is betting that they will expose core capabilities rather than absorb all of them into closed products.

The 2014 AWS analogy applies only in part. AWS won by standardizing infrastructure and allowing users to avoid directly managing underlying capacity. If context aggregation succeeds, it should similarly spare ordinary users the need to understand permission graphs and local inference. Personal data, however, is more sensitive than cloud compute, and its switching costs are higher.

04 Implications for AI Builders

The priority this week is not whether to add memory to an agent, but who should own that memory and how it should be authorized.

First, product design should distinguish among three types of state: product-level preferences, user-level personal context, and task-level transient state. Placing all of them inside a vendor-controlled session gradually turns products into immovable data islands.

Second, prioritize export and permission capabilities rather than pursuing “smarter” memory alone. At minimum, users should know which facts are stored, where they came from, which agents are reading them, and when they are being accessed. They should also have granular control and deletion options. My preliminary view is that auditability will quickly become a baseline requirement for consumer agents rather than an enterprise-only feature.

Third, the API layer should support context addressing. Model calls should do more than accept a concatenated prompt; they should also identify the source, freshness, scope, and access policy of each piece of context. Otherwise, every application integrating another agent will have to reimplement retrieval, conflict resolution, and consent.

Fourth, do not equate local execution by default with privacy or security. It may reduce centralized data exposure, but it does not address model inference, permission bypass, backups, sidecar synchronization, or prompt injection. Second’s real engineering challenge is likely securing updates to the personal model within a local background engine while allowing multiple agents to collaborate without sharing the entire context.

For independent AI-product founders, this exposes an opportunity: do not build another personal agent first; build agent infrastructure. Context sync, policy enforcement, identity, auditing, model routing, and authorization layers across MCP and A2A could become a new infrastructure category between APIs and SDKs.

Consumers of model APIs should also monitor the effect of personal context on token economics. Once context is injected continuously, long-context windows will directly increase the cost of each inference. A cheap small model with precise retrieval may be more economical than an expensive model given an entire profile. Whether prompt caching can accommodate frequently changing user state also requires separate evaluation.

05 Counterargument / Risks

I may be overestimating the viability of an independent context layer.

Google, OpenAI, Anthropic, and Microsoft all possess large volumes of user interactions, along with applications, browsers, operating systems, or enterprise distribution. They can keep personal context within their own ecosystems and expose limited capabilities through MCP and A2A. First Person Labs is not competing only with startups; it is competing with platform incumbents that control default entry points and already have switching costs.

Second could also remain a privacy-themed memory app. On-device execution can create a compelling narrative, but retention will ultimately depend on whether its context is sufficiently accurate, reduces repeated inputs, and avoids erroneous decisions at critical moments. None of the usage data disclosed by a16z establishes that these conditions have been met, and an architectural choice cannot be treated as product validation.

More seriously, the more complete the personal model becomes, the greater the payoff for attackers. Prompt injection may no longer merely cause an agent to misread a webpage; it could induce the agent to leak relationship graphs, financial information, and behavioral patterns. A local-first approach may prevent data from being centrally sold, but it does not automatically prevent local agents from abusing it.

My judgment should therefore be narrowed: First Person Labs has identified the right problem, but the winner need not be an independent context vendor. If the category succeeds, it is more likely to emerge as a middleware layer through open protocols, OEM partnerships, or model-provider partnerships, rather than requiring users to entrust all personal state to a new company.

The most important metrics to validate now are not the number of privacy features, but the context reuse rate, incorrect-memory rate, security boundaries after permission revocation, and whether users are willing to share one personal model across more than two distinct agents. If those metrics fail to materialize, “personal model sovereignty” could easily become another form of SaaS lock-in.