Giving AI coding agents a governed memory

The core failure mode of AI coding agents is familiar: they lose context between sessions, repeat mistakes, and mix temporary task state with durable knowledge. memory-os was built to fix that by giving agents a memory with rules.

Canonical truth in Git-reviewed Markdown

The system treats Git-reviewed Markdown as canonical truth. Durable knowledge lives in files that have passed review — versioned, dated, and attributed. Temporary task state is kept separate, so it can't pollute the knowledge that persists.

Retrieval and storage

GBrain provides vector and keyword search across project context. Postgres with pgvector holds structured storage and semantic retrieval. Beads handles task state, and Skills package reusable agent workflows into a repeatable operating procedure.

Evidence rules

Every fact carries evidence, a date, a source, and an owner. Setup, verify, sync, backup, and restore workflows are defined. Handoffs are verification-gated: work isn't marked done until the verification passes.

What this proves

Durable project context is an engineering problem, not a hope. Evidence-based automation discipline — where decisions trace back to retrievable artifacts — is what separates a memory system from a scrap pile.