links: private vault | no public repo or live url
one of the unglamorous parts of running multiple ai tools is deciding what they read first.
the answer has become more important than the model choice. if each tool starts with a different memory, then the system is not really shared. it is just a set of assistants with similar vibes and different blind spots.
so i tightened the entrypoint model around the vault.
AGENTS.md is the global policy layer. it covers how agents should behave
across repos and tools. CLAUDE.md is the vault operating manual. it explains
the structure, the memory workflow, the session logs, and the project map.
root-level compatibility files point into that split instead of copying it.
that sounds like bookkeeping, but it changed the feel of the whole setup. Codex, Claude Code, Cowork, and the scheduled agents can all arrive at the same place: policy first, operating context second, task context third.
the real lesson is that the vault is not just storage. it is the interface between tools. models come and go, and the orchestration layer keeps changing, but the durable state is still markdown i can read.
the agents do not own the truth. they visit it.