Codex memory between sessions

Codex builds its guidance from AGENTS.md files in a layered order: global, then project files from the root down, then subtree files where it reads. That is instructions, loaded fresh each session.

OpenAI documents a size cap on instruction files as well. Past the cap, content gets truncated without an error, so a large AGENTS.md is quietly delivering only part of what you wrote.

Inside a session, compaction rewrites older conversation. Public reports on the Codex tracker include compaction halting sessions and compaction leaving an unusually large amount of history in the active window, which pushes the same problem to the next run.

The published Codex package installs lifecycle hooks that recall relevant memory before the model runs and capture the completed turn after it, with a re-inject after compaction.

It works in the CLI and desktop surfaces, scopes to your user account, and writes to the same vault your other tools read.