Learning Brief — 2026-05-20

What we covered

• AI news: Open-Source Models Gain Traction: Rerankers and Earth Observation Get Efficiency Upgrades
• PM news: Teresa Torres Is Using AI Agents to Run Her Actual Workflow — Here's What That Means for You
• PM learning: PM Brain OS: Building a Persistent Context System for Better Decision-Making

Mental model

Externalize your PM context as persistent state, not working memory — your decision quality scales with how reliably you can access your reasoning.

Summary

Allen AI released OlmoEarth v1.1, a more efficient family of Earth observation models built on open-source foundations. These models are designed to reduce computational overhead while maintaining capability on satellite imagery tasks. Hugging Face shipped the Ettin Reranker Family, a new set of reranking models for retrieval-augmented generation pipelines. These are purpose-built to improve ranking quality in RAG workflows without the latency or cost of full LLM inference.

Teresa Torres just published something that feels like a glimpse into how senior PMs might actually work in 2026. She's running a team of Claude agents that operate while she's away from her computer. Her podcast-manager agent preps interview notes. Her task-management agent surfaces priorities each morning. The agents write to her systems, read her calendar, and operate with enough autonomy that she's not micromanaging them — she's just reviewing what they've done.

Here's why this matters beyond the novelty: this is a real PM testing the boundary between delegation and automation at scale. For years we've talked about AI assistants as tools you query. Torres is treating them as team members with standing instructions and persistent context. That's a different mental model entirely.

The PM angle is structural. As you move toward GPM or Chief Product roles, your bottleneck stops being execution and starts being decision quality and coverage. You can't read every customer interview. You can't synthesize every metric. You can't prep for every meeting. Right now, you hire people or you skip things. Torres is showing that there's a middle path: agents with clear scope, persistent memory of your systems, and enough autonomy to surface what actually needs your attention.

The technical implementation matters less than the workflow question it raises: what tasks in your current role are actually just "read context, apply judgment, write output"? Because those are exactly the tasks that agents can handle if you set them up right. Interview synthesis. Competitive brief updates. Metrics dashboards. User feedback categorization. The question isn't whether AI can do these — it's whether you're structured to let it.

If you're building toward a GPM role, start thinking about which parts of your decision-making process could run on standing instructions rather than your direct attention.

Here's the thing: most PMs operate with working memory. You're in a meeting, you make a call based on what you remember from last week's research or that Slack thread from two months ago. And most of the time, you're working with incomplete or stale context. Pawel Huryn just shipped something that reframes how you can actually solve this — not with a fancy tool, but with a system that treats your knowledge like a persistent operating system.

The core insight is this: if you externalize your PM context — your frameworks, your user research, your decision history, your strategic assumptions — into a structured markdown folder, and you make it available to Claude before every conversation, you get something powerful. You're not starting from zero each time. Your AI assistant reads your PM brain before answering, then writes back to it. Every Friday, you sweep and consolidate.

What that means in practice is you stop repeating yourself. You stop re-learning the same user insight three times in different contexts. When you're in a prioritization meeting and someone asks "why did we kill that feature?", you don't rely on memory — your system knows. When you're briefing a new stakeholder, Claude has your full context and can explain your thinking in your voice, not generic PM-speak.

The move here is treating this like a second brain for your role, not a productivity hack. This isn't about automating tasks — it's about building institutional memory that travels with you. Think of it like version control for your thinking. You're creating a persistent state of your product strategy, your user understanding, your decision rationale. That state gets richer every week.

Why this matters at your level: as you move into Senior PM or GPM roles, you're managing more complexity, more stakeholders, more history. The PMs who win aren't the ones with the best memory — they're the ones who can access their context consistently and help their teams do the same. This system lets you scale your thinking. You're not bottlenecked by what you can hold in your head.

The concrete output: you get 99.5% consistency on PM decisions because your reasoning is documented and accessible. That's not a vanity metric — that's alignment. That's faster onboarding for new team members. That's being able to explain your strategy six months later without sounding like you're making it up.

This week, audit one major decision you made in the last month. Write down the actual reasoning — not the polished version, the real one. Where did it come from? What context did you use? What did you assume? That's your starting point for building this system.