Skip to content
Mike Reams
← Work

PRJ · 2024

AI-Maintained Architecture Second-Brain Repository

A governed, AI-maintained knowledge base for an EA practice: raw sources in, cross-linked wiki out, under a written contract.

Scattered EA docs to AI-maintained graph

System diagram: raw sources flow to operator and architect agent tiers, which maintain a curated wiki that passes a reviewer before anything is published.System diagram: raw sources flow to operator and architect agent tiers, which maintain a curated wiki that passes a reviewer before anything is published.

What it is

An AI-maintained knowledge base for an enterprise architecture practice. Source material — inventory exports, meeting notes, vendor documents, research — goes into an immutable folder. An AI agent reads it, writes and cross-links governed wiki pages for applications, capabilities, decisions, standards and vendors, and keeps the whole graph consistent under a written operating contract. I curate, ask questions and direct.

The problem

Architecture knowledge lives everywhere at once: decisions in slide decks, the application inventory in the CMDB, reasoning in people's heads. Every new question starts with an afternoon of digging, and every answer goes stale the moment it is written down somewhere else.

What I built

  • Three layers and a contract. Immutable evidence, an AI-owned wiki, and an operating contract that defines page types, rules and rituals.
  • Resume state. A small set of files and fixed open and close rituals so work survives across sessions, days and machines.
  • Two agent tiers. A strong model for schema and judgment, a cheaper model running a registry of bounded, repeatable skills.
  • A skeptical reviewer. An independent pass that assumes a change is wrong until its checks prove otherwise, and never fixes its own findings.
  • Validators and gates. Integrity, provenance, schema and link checks that run before anything is trusted or published.
  • A two-lens capability model. Every application mapped to the business outcomes it serves and to exactly one technical class.

What came out of it

Answers that used to take an afternoon now come back in minutes, with citations to the pages they rest on. Figures live in one canonical place, so a change propagates instead of leaving stale copies behind. And the practice keeps a durable, reviewable record of why decisions were made, not just what they were.

A starter version is public: ea-brain on GitHub, MIT licensed, with the contract, resume state, rules and skills and no organization's data in it. The write-ups below cover the patterns one at a time.