AI writes the code. Your architecture still decides what ships.
Frontend surfaces may request analytics through the backend service. They must not instantiate data-platform clients directly.
// frontend/checkout.ts
import { BigQuery }
from '@google-cloud/bigquery';
export async function
trackCheckout(event) {
const bq = new BigQuery();
return bq.insert(event);
}
ADR-017 forbids a data-platform client in the frontend checkout service.
Had it shipped Warehouse credentials bundled into the browser, and a customer-data path nobody reviewed.
Clone the open-source repository and watch an architectural violation get caught before generation — no API key required.
$ git clone https://github.com/MnemeHQ/mneme
$ cd mneme
$ pip install -e .
$ python examples/demo-adr-import.py
WARN [ADR-005] constraint "no MnemeHQ" -- trigger: mnemehq
Brand vs Package Namespace Enforcement
Result: WARNInstall the owned package, initialize the decision corpus, add one approved decision, and check the first proposed change.
$ pip install mneme-hq
$ mneme init
Created .mneme/project_memory.json
$ mneme check \
--memory .mneme/project_memory.json \
--input proposed-change.diff \
--query "checkout analytics"
FAIL · ADR-017 · exit 2
Result: FAILA Mneme pilot turns one real set of architectural decisions into an enforceable workflow across the agents and repositories your team already uses. Start with one boundary, one repository, and evidence everyone can inspect.
People decide. Mneme makes the decision operational. Evidence proves it held.
Mneme is open-source architectural drift prevention for the AI SDLC. It turns approved ADRs, engineering standards, project context, and prior decisions into deterministic guide, warn, or block verdicts before incompatible changes land, with every verdict traced to its source.
Mneme can apply relevant decisions before an agent writes code, during file changes through supported hooks, and in CI before a change merges. Each guide, warn, or block verdict identifies the governing source and permitted next action.
Context systems help agents understand. Auditing tools show how agents behaved. Code review evaluates completed output. Mneme enforces the approved engineering intent the agent must follow, while linking every verdict to its source.
