Introducing ea-brain: A Starter for an AI-Maintained Architecture Knowledge Base

This series has been about running an AI-maintained knowledge base for an enterprise architecture practice: resume state, a skeptical reviewer, two model tiers, the two-lens capability model.

Each post ended with something to copy. ea-brain collects those pieces in one repository: github.com/solventarchitect/ea-brain, MIT licensed. Click Use this template and it is yours.

What is in it

• The operating contract. CLAUDE.md: two layers (raw sources in, a wiki the agent owns), six rules, and the session open and close rituals.
• Resume state. wiki/NOW.md — active threads, tripwires, pending ratification and a ruled-out list — with an index, an inbox and an append-only log.
• Seven rules, written to paste in whole. Verify before you assert, pushed is not deployed, guarded writes, ServiceNow Table API behaviors, the two-lens capability rule, the record rule and contract reconciliation.
• A reviewer and an operator. The skeptical-reviewer skill, the operator agent's instructions, and a template for a bounded, repeatable skill.
• Two examples. A disposition record with evidence and a review date, and the shape of a golden-questions eval set.

The README maps every file to the post that explains why it exists.

ea-brain/
├── CLAUDE.md            operating contract
├── wiki/
│   ├── NOW.md           resume state
│   ├── index.md         catalog of every page
│   ├── inbox.md         open questions
│   └── log/             append-only, one file per month
├── raw/                 your sources (git-ignored)
├── rules/               seven rules, paste in whole
├── agents/operator.md   the bounded operator
├── skills/              skeptical-reviewer, _template
└── examples/            disposition record, golden questions

What it deliberately is not

A starter, not a product. There are no scripts, validators, metamodel or integrations. That is on purpose: the patterns travel between organizations, and the tooling does not — it gets built around your own systems of record. Nothing in the repository comes from any organization's data, and raw/ is git-ignored by default because sources usually are sensitive.

Getting started

1. Make your copy. Use this template on GitHub, or clone it.
2. Name your practice. Replace <team/practice> in CLAUDE.md, and change anything else that does not fit.
3. Add one source. Drop an export, meeting notes or a vendor notice into raw/.
4. Open an agent session in the folder and ask it to ingest the source. It reads CLAUDE.md, then NOW.md, and writes pages under wiki/.
5. Close every session with the close ritual. The files are the memory.

How to grow it

Start with the contract and NOW.md alone. Add a rule the second time you catch the same mistake. Add the reviewer before the agent writes anything someone else will read. Add the operator and a cheaper model tier when routine work starts crowding out judgment.

What I would love back

Issues on the repository describing where the contract was unclear, where the agent went wrong, and which rule you had to add. The pattern got good because every failure became a rule. It will get better the same way.