What they did
Terui wanted one memory that any agent could use, instead of each AI service keeping its own. Since 5 October 2026 they have kept memory and rules in plain Markdown files in a Git repository, with AGENTS.md as the source of truth and CLAUDE.md importing it for Claude Code. Numbered folders separate rules, private data kept outside Git, facts about Terui, output, research and todos, and the agent reads only what a task needs. It answers from the files, writes records, adds calendar events and drafts emails; sending and deleting stay with Terui. Three scheduled jobs run: a morning summary, daily AI and IT news, and a Sunday weekly review.
How it went
Terui has used it every day for about six weeks, with around 680 commits. The agent kept making plausible mistakes, and each became a written rule: run date before saying anything about time, follow a fixed order of sources before concluding a feature does not exist, run diagnostics before naming the cause of a tool error, and label every statement as fact, guess or unconfirmed. Overwriting a note body with a script shrank its font, so the rule now is to create a new note instead.
Worth knowing
The fixes are procedures, not requests to be careful, and they survive a change of model because they live in the files. A redacted version of the design is public on GitHub. The post does not name the calendar or mail apps it drafts into.