I gave my AI agent's memory a CI/CD pipeline

Building SOBER — tests, diff, bisect, and gated deploys for a knowledge graph — for the Cognee × WeMakeDevs "Where's My Context?" hackathon.

TL;DR — Your agent's memory is production infrastructure with no tests, no diff, no rollback. SOBER is a brain CLI + GitHub Action that wraps a Cognee knowledge graph in real CI/CD: forget-regression tests that prove a retracted secret stays gone, git bisect for a poisoned graph, and a nightly improve() that opens its own pull request behind a green eval gate. Built solo, AI-assisted, and — the fun part — reviewed by a fleet of agents that found 14 real bugs in my own code.

Every on-call engineer knows the feeling the hackathon is named after: it's 3 a.m., something is broken, and nobody remembers how it got fixed last time.

We solved that for code decades ago — version control, tests, code review, rollback, canary deploys. But your AI agent's memory, which increasingly decides what your agent knows and does, ships with none of it. It mutates in place, silently. There is:

So I built SOBER — CI/CD for Agent Brains. Not memory for DevOps. DevOps for memory.