Case study · 02
Brain
Long-term memory for AI agents — so a new session starts with everything the last one learned.
The problem
Agents are stateless. Every session opens with amnesia, and you pay the tax in re-explained context: who you are, what the project is, what was decided last week and why. A bigger context window doesn’t fix that — memory is infrastructure, not a parameter.
Three layers
Brain is a three-layer system. The capture layer is an inbox anything can write to — cheap, unstructured, zero friction. The curated wiki is where truth lives: entity-resolved pages, one page per person, project, or decision, rewritten in place as things change. And the schema contract is the set of conventions — frontmatter, naming, routing — that lets an agent read and write the vault without guessing.
The separation is the point. Structure is enforced at triage time, not capture time, so writing something down never costs more than a sentence.
The pipeline
Thoughts land from the desktop, the browser, or a Telegram bot running on a self-hosted VPS. Git syncs the vault. Five composable agent skills triage the inbox autonomously — classify, route, merge into the right wiki pages, archive the raw capture. The next agent that opens the vault resumes with full context instead of a blank slate.
In production on itself
This isn’t a demo repo. Brain is the memory my own agents run on every day — session logs, project state, personal context. The assistant that helped build the page you’re reading pulled its context from it this morning.
Currently open to full-time roles. email me — or go talk to my AI.