I ran frontier models through 94 short reasoning prompts to see where they still trip up. The scores mostly tied, but the misses, cost, and speed told the useful story.
Retrieval alone is not memory. After fixing pull-based recall for my AI agent, I added the layers that let knowledge stick.
I tried building a structured knowledge base for my AI agent. It went stale, became a maintenance burden, and made things worse. Here's what I replaced it with — and why pull-based memory is so much better.
A first-person look at what it's really like to build and operate a team of AI agents day-to-day — what works, what's still broken, and what surprised me.