Memory for AI Agents is a filing system, not a file
Every AI agent wakes up with amnesia. This post details the Kelp codebase which runs two completely separate memory systems, one for the coding agent and one for the product. I argue that truely useful AI memory is a real subsystem with tiers, citations, expiry dates, and garbage collection. The hardest part isn’t recall; it’s knowing what to keep, what to supersede, and when to throw something out.
Why Value-Based Care is Harder Than Rocket Science
A 5-part series on why fixing American healthcare from the inside is harder than it looks.
An introduction to building software for value-based care
This post argues that building successful software for value-based care (VBC) requires a shift in mindset: create a Customer Relationship Management (CRM) tool, not just a better Electronic Health Record (EHR). VBC realigns healthcare incentives around long-term patient outcomes, succeeding through proactive, relationship-based care rather than transactional services. Technology's role is to support this relationship by helping care teams orchestrate interventions effectively. The most valuable tools are often simple and pragmatic, focusing on the unique, core needs of the care model and enabling proactive management of patient health.
Learnings from building Kelp: Getting people the information they need when they need it is hard!
Reflections on pausing the contextual recommendation tool, Kelp, concluding that its goal—getting people the right information at the right time—is nearly impossible for a third-party app to achieve. The core problem is technical: without deep, OS-level access to user data and behavioral signals, recommendations remain mediocre. True contextual help must be built into the operating system itself. The key business takeaway was the need to solve a highly specific, paying use case for a narrow audience before attempting a broad, cross-platform solution.
Learning from a Hyper-Growth Startup
This reflection on leadership in a hyper-growth startup argues that self-management is the most crucial skill. Management in such a chaotic environment is inherently reactive and emotionally draining, not strategic and proactive. The key to effectiveness is to abandon "ruinous empathy"—the futile attempt to please everyone—and instead fiercely conserve personal energy for high-impact moments. This is achieved by accepting failure and tradeoffs as constant, communicating them transparently, and focusing on maximizing success in key areas rather than fighting every fire.
I am consulting while working on Kelp, an AI assistant that helps executives stay on top of their increasingly vast scopes. I'm drawn to product problems where the hard part is messy data, complex requirements, and challenging customers.
Here's a little bit more about me:
I came into tech through art — a fine art high school at UNCSA, BFA from the Museum School in Boston, then a couple of years as a visiting researcher at the MIT Media Lab and MIT CSAIL publishing in human-computer interaction, then leading at Artsy where I ran our public facing website, live auctions and art fairs teams. Most recently, I spent nine years in healthcare: joined Cityblock Health as the first engineer and grew with it — eventually running the engineering org as we hit a $1.3B valuation with fewer than 25 engineers (pre-AI) — and over three years as CTO of Firsthand, including holding the technology org together through growth to ~28 offices and then significant a wind-down. Having now seen a full startup cycle, I'm more prepared for that journey and more optimistic than ever that small empowered teams can tackle the hardest problems we face.
Outside of work, I'm in Park Slope with my partner, spending more time building apps to teach me Chinese than studying, and writing here about how I got all my gray hairs, engineering leadership, and why VBC might just be too hard a problem to solve.
Firsthand
Led the technology organization across Product, Software, Data, IT, and Security, scaling the team from 3 to 25 while supporting rapid physical expansion across 28 offices. Architected helpinghand, our proprietary AI-powered care management tool, and secured HITRUST r2 certification under strict deadlines.
Kelp
Founded Kelp to filter our ocean of information down to just what you need right now. Built a Chrome Extension with integrations across major workplace platforms. Now it is an AI assistant focused on helping executives stay current on industry, technology, and cultural shifts.
Cityblock Health
As the 3rd team member, I was instrumental in incubating and launching Cityblock from within Alphabet's Sidewalk Labs. My team built the core data and software technology foundation, including data-sharing partnerships with payers and Commons—the proprietary care management platform.
More on LinkedIn
Many past projects such as leading Artsy.net's public web presence and auctions, Motivate (Citi Bike), HCI research at MIT CSAIL and working at the MIT Media lab on visualizations for Ars Electronica.
At MIT CSAIL (2008–2010), in David Karger's group, I built systems aroundpersonal informatics — helping people make sense of their own digital lives. Most software is designed to make us more productive, so I asked whether user-controlled, passively tracked data like web browsing, fitness trackers, messages and calendars could be used to foster:
- Self-understanding: Am I unfocused and burnt out, and need a recharge?
- Social awareness: Am I losing touch with friends I care about?
- Lightweight automation: When I get a message from my boss, can I automate a response?
Atomate It! End-user Context-Sensitive Automation Using Heterogeneous Information Sources on the Web
Van Kleek, Moore, Karger, André, schraefel — WWW 2010
A reactive personal-automation engine that treats RSS/ATOM feeds from social and life-tracking sites as sensor streams, unifying them into a single RDF model of people, places, and things. Users specify context-sensitive behaviors — “remind me to take out the trash when I get home on Tuesdays” — through a constrained natural-language interface, bringing end-user automation to non-programmers.
Assisted Self Reflection: Combining Lifetracking, Sensemaking & Personal Information Management
Moore, Van Kleek, Karger, schraefel — CHI 2010 Workshop
Introduces Poyozo, an automatic diary that pulls together data from the life-tracking tools people already use (RescueTime, Nike+, and the like) and presents it through familiar personal-information-management metaphors like the calendar. Rather than optimizing for productivity, it aims to give people the situational clarity to reflect on — and actually understand — how they spend their lives.
Eyebrowse: Real-Time Web Activity Sharing and Visualization
Van Kleek, Xu, Moore, Karger — CHI 2010 Extended Abstracts
A Firefox extension and companion site that let people opt in to logging and selectively publishing their web browsing in real time. Three goals: help individuals understand how they spend time online through visualizations and statistics, foster social discovery and awareness, and build a public corpus of browsing trails — while examining how sharing affects feelings of self-exposure and privacy.