The cost of memory-as-files isn't writing them. It's that the agent will cheerfully claim it updated something and not actually do it, or write a one-line stub that satisfies the spec but loses the original signal. Without a verification layer, the vault accumulates plausible-looking entries that quietly drift from reality.

What ended up working for me was treating the agent's self-reported summary as a wish, not a fact. A separate process diffs the actual file system against the claimed changes and flags mismatches.

After a few cycles, the agent gets calibrated and stops claiming things that don't survive a file check. That has the side benefit of making the diff review itself much higher signal: most of what shows up is real.

The split I'd make early is per-agent instructions vs. cross-thread shared notes.

They sound like the same artifact, but “what this agent should always do” and “what sibling work just learned” age very differently. Mixing them means the wisdom gets stale together.

Give each Codex an AgentName and ask them to mark their PR/issue/comments with those. Have one or two "managers" that manage PRs and overall project direction. I write the project directions and make long lasting issues. Each Codex session has an almost unachievable `/goal` but they are asked to achieve the goal by landing changes in `main` via PRs

I am running about 14 Codex sessions on 4 machines right now for about two weeks since OpenAI 10x'ed my 20x account and I simply can not run out of tokens fast enough.

Side note: I have multiple Claude accounts too but the new Claude Code `/goal` command is seriously broken. It waits long pauses between iterations and sometimes prematurely stops.