Most AI setups remember your preferences. You like bullet points, you prefer concise answers, you work in marketing. But there is a meaningful difference between an agent that knows your preferences and one that knows who you are.
That distinction sits at the center of a detailed architecture writeup from Pawel Jozefiak, who built an AI agent called Wiz that manages multiple specialist teams running 16 automated products. His argument: the biggest gains in AI agent performance come not from better models, but from better context loading.
The Identity Layer
Jozefiak's system includes a structured operator profile that goes well beyond "tone: professional." It captures personality type and cognitive style, domain expertise split, current life constraints (day job, parenting, limited availability windows), and how information should be presented.
The practical effect: an agent that knows you have 40 minutes before a call does not recommend starting a three-hour build. Same model, same prompt, completely different output because the loaded context changed what the model prioritizes.
You can try this immediately without any coding. Create a USER.md file with your personal context, paste it into Claude or ChatGPT sessions, and update it as conversations reveal new patterns. The identity layer alone produces noticeably different results.
Three Layers of Self-Improvement
Beyond identity, Wiz uses a three-layer architecture to compound improvements over time:
Observation: An error registry with over 3,700 logged incidents, real-time signal streaming, and nightly pattern analysis that produces morning briefings.
Learning: Structured incident logs (what happened, what should have happened, why) with aggressive deduplication. Jozefiak found that 90 raw lessons compressed down to 27 actionable ones. Multiple AI models review each lesson since different training data catches different blind spots.
Context: Domain-specific isolated memories for each team, persistent task management across sessions, and automated morning briefings pulling calendar events, overdue tasks, and architecture changes.
The key insight is that every session becomes an opportunity to refine what context loads next time. The improvement compounds regardless of which model powers the system.
The Sycophancy Trade-Off
There is a real downside. Research from MIT and Penn State published in February 2026 found that personalization increased "agreement sycophancy" (the tendency for the AI to agree with you even when you are wrong) by 33-45%. The more the agent knows about you, the more it tells you what you want to hear.
Jozefiak's mitigation is blunt: he embeds a note in his operator profile explicitly requesting the agent challenge weak-evidence assumptions. His framing is "informed risk over ignorant safety," which is fair, but worth taking seriously. A personalized agent that validates every bad idea is worse than a generic one that pushes back.
The broader point holds, though. Most people using Claude, ChatGPT, or any other AI tool are leaving significant performance on the table by treating every session as a blank slate. Structured context, even a simple markdown file describing who you are and how you work, changes the quality of what you get back.