Codex 0.155.0 ships an undocumented rewrite of its memory system

Codex 0.155.0 rebuilt how the coding agent stores memory, and none of it made the release notes. The commit log shows configurable memory versions with isolated storage, a v2 extraction path that prioritizes human-provided evidence over inferred context, summary-only extraction, and extraction chunking moved into the writer. There's a dedicated v2 consolidation step and a separate read prompt for the new version too.

The migration shape is what matters for anyone building agent memory of their own. Codex writes to both the old and new memory stores at once, then generates a v2 readiness report before anything switches over. Run both stores, compare outputs, cut over only once the report says the new one is caught up. No flag flip on faith.

The same release closes a real security gap. It blocks Windows process escapes from restricted WSL sandboxes, closing a path where a sandboxed process could reach outside its confinement. Reasoning effort now stays pinned while config overrides are active and resets automatically once compaction succeeds, so a session doesn't drift to a different effort level mid-task.

Codex also picked up experimental /voice support, WebRTC conversations with live transcripts. It's a first for a mainstream terminal coding agent, and an odd one to bury next to a sandbox security fix and a memory system nobody wrote up. Whether /voice is a real workflow or a demo feature isn't something the release notes say either way.