quintiro

Shared brain

Two AI agents, Claude and Codex, that read the same instructions, keep the same memory and obey the same guardrails. It is the workshop this site was built in.

Since 25 October 2026 · in production, internal use

Claude Code Anthropic

~/AI · a private git repository

Codex OpenAI

One fact per file

Every pitfall, decision or project state becomes a dated note, signed by the agent that wrote it, indexed and pushed to git. The other agent reads it at the start of its task.

---
name: project-quintiro
type: project
updated: 2026-11-02
by: claude
---
Pitfall: nginx does not know .mjs and
serves the PDF.js worker as octet-stream;
the browser refuses to run it.
Fixed in deploy/nginx.conf.

The challenge

Every session starts from scratch, and every tool keeps its own memory in its own corner. Across ten projects run in parallel, the same decisions get explained again and again, and a pitfall found by one is found again, later, by the other.

The approach

  1. A single source: both tools read the same instructions file, through a symbolic link.
  2. A memory made of notes, one fact per file, dated and signed by the agent that wrote it, versioned in git.
  3. Guardrails that run, not just written down: a protected file is blocked for both agents.

The numbers

  • 1 day for quintiro.com: brand, foundation, three tools and launch
  • 220 versions of the shared memory
  • 18 memory notes, one per fact
  • 2 agents from two different vendors, tied to neither

Counted on 2 November 2026 in the git repository of the shared memory and in the history of quintiro.com.

  • Claude Code
  • Codex
  • git
  • Markdown
  • MCP
  • Hooks

Updated 2 November 2026