From personal pain to product
The original question was simple: why is it so hard to actually follow through on the things you care about? Not for lack of intent — for lack of a system that understands you, adapts to your energy, and keeps you moving without turning into yet another productivity app that demands the executive function it was supposed to help you build.
Sai and Shivanshu — a machine learning engineer and a product leader — had both lived this. They’d also watched friends and family struggle with ADHD in a world where the available tools were either pharmaceutical, expensive coaching, or to-do lists built for people who didn’t have the problem.
They interviewed approximately 35 people and 10 families across three segments — tech professionals with symptomatic ADHD, busy families, and university students — using structured usability tests and open-ended user research sessions. That research shaped every product decision: what the AI said first, how it framed goals, when it checked in. The coaching model drew on CBT frameworks, NIH-sponsored ADHD motivation research from Duke University, and Dr. Andrew Huberman’s work on goal-setting neuroscience.
The application reached the top 10% of Y Combinator applicants in 2023 — external validation of both the technical depth and the relevance of the problem.
The pivot: finding the real infrastructure problem
Building a personalized AI coaching agent forced a hard realization: LLMs are stateless, and the ecosystem had no good solution for it. Increasing context window size created latency and cost problems. RAG over interaction logs was a partial fix at best — it lacked human-like cognition and degraded over time. Topic summaries lost rich detail. Every approach had fundamental trade-offs, and there was no off-the-shelf memory layer to reach for.
This was 2023 — before MemGPT, before mem0, before the broader AI industry coalesced around “agent memory” as a distinct infrastructure problem. The team built their own solution and realized the problem was universal: nearly every AI application developer trying to build personalized, stateful agents was hitting the same wall. Pinecone’s sales team reached out independently, describing the exact gap their enterprise customers were trying to solve. This wasn’t a niche problem.
The pivot: build that memory infrastructure as a developer-first API. The closest legacy analogy was a CDP — but AI-native, covering short-term, long-term, episodic, factual, and procedural memory types through a single integration point.