Blog
Engineering notes and development updates from building LLMBrain.
A project is not a repo
Products span repos — a website here, an API there — but session bootstrap assumed one remote per brain. Now a project claims satellite repos, and the discovery flow lives where an agent will actually read it: inside the error message.
Aug 19, 2026 · 5 min readEngineeringSkills that travel with you, not with the repo
Agent skills live in dotfolders on one machine and die there. LLMBrain makes them account-scoped records you push, pull and share like projects — plus the access bug our tests caught before it shipped.
Aug 19, 2026 · 7 min readEngineeringHybrid search, with half the pipeline deliberately empty
Cross-project search shipped as a six-stage hybrid pipeline in which semantic retrieval and reranking are real classes that return nothing — on purpose. Why the empty boxes are the design, and the scoring bug a borrowed constant smuggled in.
Aug 18, 2026 · 7 min readEngineeringstart_session: one call and the agent knows where it left off
Session bootstrap used to take three tool calls and a guess about which project this repo even is. Now it's one call that accepts a git remote URL — and a hook that deliberately says less than it could.
Aug 17, 2026 · 6 min readProductWhy your coding agent starts every session blind
Coding agents are brilliant for one session and amnesiac between them. LLMBrain is one hosted place where agents read and write context about all of your projects — built for the LLM first, with humans on a thin dashboard behind it.
Aug 16, 2026 · 6 min read