Codex 0.120 adds per-project memory extensions and Realtime V2 streaming

• Codex CLI 0.120.0 shipped Realtime V2 background progress streaming and queued follow-up responses, while the release highlight tweet surfaced the update to the broader Codex audience.
• Kevin Kern's walkthrough showed the most interesting addition: per-project instructions.md files under memories_extensions/<project>, which steer what the memory consolidation agent should and should not retain.
• OpenAI's own customization docs still position AGENTS.md as the always-on repo rulebook, and Kern's example makes the split explicit by treating memory extensions as guidance for what Codex learns across sessions.
• TestingCatalog's Scratchpad clip exposed an unreleased TODO-style view that fans out multiple Codex chats in parallel, and TestingCatalog's write-up ties it to a broader unified Codex app.
• Separate app findings from chetaslua's automation screenshot and a follow-up thread point to Codex growing beyond a chat box, with heartbeat automations, thread routing, and connected-app context in the same surface.

You can read the full 0.120.0 changelog or the matching GitHub release notes, compare that with OpenAI's customization model, and then jump to TestingCatalog's report for the unreleased app pieces. The oddest reveal is that the public release is mostly CLI plumbing, while the leaked app UI already shows automations, task lists, and cross-thread routing.

The new memories_extensions/<project>/instructions.md file is a second control plane for Codex. According to Kern's example, it tells the memory consolidation agent what patterns to promote, what to ignore, how project-local the memory should stay, and which recurring workflows count as durable knowledge.

That makes it different from AGENTS.md, which OpenAI's customization docs describe as durable project guidance that applies before the agent starts work. Kern frames the split cleanly: AGENTS.md is for rules Codex should obey immediately in every session, while the memory extension file is for rules about what it should learn from repeated past work.

The screenshot attached to that post also shows the granularity OpenAI is aiming for:

• identify which rollouts belong to a project
• promote stable architecture and workflow patterns
• keep one-off details out of memory
• shape the tags and output format of generated memory artifacts

The official changelog and GitHub release make 0.120 look like a quality-of-life release for people already living in Codex. The headline change is Realtime V2 streaming background agent progress before the job finishes, plus queued follow-up responses that wait until the active run completes.

The rest of the release is a tight batch of operator-facing fixes:

• easier-to-scan hook activity in the TUI
• renamed thread titles in custom status lines
• more precise typed tool results via MCP outputSchema
• a new SessionStart distinction for /clear
• sandbox fixes for Windows and symlinked writable roots

The unreleased piece is Scratchpad. TestingCatalog's video shows a TODO list spawning three Codex chats in parallel, then collecting them into a dedicated Scratchpad view.

TestingCatalog's article says the feature is meant to support a broader Codex app where users can trigger a wider range of tasks from one place. That lines up with the UI in the clip, which treats parallel task execution as a first-class workflow instead of a hack built from separate threads.

The app evidence points in the same direction. In chetaslua's screenshot, a Codex chat can schedule an every-30-minute automation called "The Next-Move Scout," suggesting a persistent agent loop rather than a one-shot prompt.

A second post from the same account adds more scaffolding around that idea. The screenshot text describes exactly three fresh suggestions, each carrying a title, description, prompt, threadAction, and appIds, plus logic for deciding whether work belongs in a new thread or an existing one.

Other sightings fill in the desktop surface. dkundel's demo shows the macOS plugin building a menubar app from one prompt, and a later post notes that the Codex app already exposes PR status, CI checks, comments, merge state, and review state inside the client.