Memory: Your Preferences Persist Across Sessions
Any developer who has used AI coding tools for more than a week has lived through the same friction: you explain your preferences once — TypeScript strict mode, named exports, your commit message format, which function in your codebase should never be touched — and the model forgets everything at session end. Tomorrow you explain it again. The repetition is not catastrophic, but it accumulates.
Codex memory is in preview as of April 16, but the capability addresses this directly. Codex now stores context from past interactions: preferences you have stated, corrections you have made when outputs were wrong, patterns it has observed across your sessions, and project-specific information it had to gather manually in previous sessions. When you start a new session, that knowledge is already loaded.
The practical effect shows up as faster first-pass quality on repeated task types. When you ask Codex to scaffold a new component, it already knows your file organization convention. When you ask it to write a test, it knows you use Vitest. When you ask for a commit message, it uses your team’s format. None of these are dramatic wins in a single session. Compounded over weeks of daily use, they shift the experience from “capable but amnesiac” to something that actually learns your workflow.
Memory can be reviewed and edited in Codex Settings. Retention follows OpenAI’s standard data handling policies. For teams working on sensitive codebases, understanding the retention scope before relying on it for confidential project context is worth doing during the preview period.
Image Generation with GPT-Image-1.5
Codex can now generate images directly inside a development workflow using GPT-Image-1.5, OpenAI’s latest image model. The integration is most useful when code and visuals are tightly coupled rather than as a general-purpose image generator.
Three use cases stand out as genuinely workflow-changing. First, UI mockup generation: describe a component and Codex produces a visual mockup alongside the implementation code, so design decisions happen before you write a line rather than after. Second, game asset creation: for game projects Codex is actively building, it can generate sprite, texture, or concept art assets without breaking out of the coding workflow. Third, frontend design exploration: when a feature’s visual direction is ambiguous, Codex can generate layout concepts as images before any code is written, surfacing design tradeoffs earlier in the process when they are cheaper to resolve.
For backend, API, or systems work, image generation rarely belongs in the workflow at all, and the integration stays out of the way. It surfaces when you are working on visual components and disappears when you are not.
90+ New Plugins: Three Layers of Capability
The Codex plugin library grew from approximately 40 to 90+ with this update. OpenAI organized the expansion into three distinct categories, each extending Codex’s capabilities in a different way.
Skills are task-specific capability extensions — specialized code review patterns, domain-specific problem-solving, documentation generators, and testing framework integrations that go beyond the base model’s default behaviors. They add what the model can do, not just what it can connect to.
App integrations are direct connections to external services. The new additions include several that meaningfully change how development teams can work:
- Atlassian Rovo — Codex can read, create, and update JIRA issues directly. A ticket describing a bug can become the input to a fix: Codex reads the ticket, implements the change, and updates the issue status without any copy-pasting between tools.
- GitLab Issues — equivalent JIRA-style integration for GitLab-based teams.
- CircleCI — Codex can inspect failed CI pipeline runs, read the failure output, and diagnose or fix the underlying code issue without you acting as the intermediary between the CI log and the editor.
- CodeRabbit — integration with the AI code review platform, bringing automated review feedback into the Codex workflow rather than a separate dashboard.
- Microsoft Suite — Codex can read from and write to Microsoft 365 documents. For enterprise teams whose project specifications and documentation live in Word, OneNote, or SharePoint, this makes Codex able to work directly from those sources rather than requiring you to extract and paste content manually.
- Neon by Databricks — serverless Postgres integration for database-heavy development workflows.
- Render — deployment platform integration for triggering and monitoring deployments directly from Codex.
- Remotion — programmatic video generation library integration, relevant for teams building video automation or data visualization tooling.
- Superpowers — an AI-powered workflow enhancement layer that expands what individual plugin calls can accomplish.
MCP servers are the third category and arguably the most extensible one. Model Context Protocol — the open standard originally developed by Anthropic — has become the common interface for connecting AI models to external tools and data sources. Any MCP-compatible server can be configured as a Codex plugin. The 90+ official plugins are a floor, not a ceiling: organizations running custom MCP servers for internal tools, proprietary databases, or specialized workflows can wire them into Codex using the same mechanism as the official integrations.
Remote Devboxes, Multiple Terminals, and Rich File Previews
Three additions in this update are quieter but worth flagging for development team workflows.
Remote devbox support via SSH is in alpha. Codex can now connect to remote development environments over SSH, which means its agent capabilities are not limited to your local machine. For teams that standardize on cloud-based dev environments — GitHub Codespaces, Gitpod, Coder, or internal dev infrastructure — this means Codex can operate in the environment where your actual code lives, not a local clone of it. Computer use and terminal capabilities extend into the remote session.
Multiple terminal tabs let you run parallel shell sessions inside Codex. Previously, a single terminal context serialized tasks that could logically run in parallel. With multiple tabs, you can have a dev server running in one terminal, a test runner in another, and a build process in a third — with Codex monitoring and interacting with all of them simultaneously rather than one at a time.
Rich file previews added rendering support for PDFs, spreadsheets, slide decks, and documents directly in the Codex sidebar. When your project context includes a product spec in a PDF, a data model in a spreadsheet, or a design brief in a slide deck, Codex can read and reason about that document directly. You no longer need to extract and paste the relevant content manually before asking Codex to work from it.
What This Update Signals About Where AI Coding Is Heading
Read individually, each feature in this update is useful. Read together, they describe a specific architectural direction: Codex is being built toward an agent that can receive a task and complete the full implementation cycle with minimal human coordination work at each step.
Consider what that cycle looks like today: a developer reads a JIRA ticket, opens a file in an editor, writes or modifies code, checks it in a browser or simulator, fixes issues, opens a terminal to run tests, commits the change, opens a PR, and then cycles back through review comments before the work ships. Each of those steps required the developer to act as the coordinator: reading one tool, translating the output, moving it to the next tool, and repeating.
After April 16, Codex can read the JIRA ticket directly, write and modify code, open the browser to check the result, see and click through a simulator, run tests in a terminal tab, and address PR review comments — all within one agent session. The developer still drives the work. But the number of steps that require the developer to manually bridge between tools is shrinking with each update.
The “(almost)” in the product name is honest. Computer use is macOS-only and geographically restricted. The browser cannot handle authenticated sessions. Memory is in preview with retention limits. Many of the new plugins are early-stage integrations. These are not minor caveats — they represent real gaps in daily professional workflows. But they are the expected state of an ambitious product moving fast, not signs of fundamental limitations.
For developers evaluating their AI coding tool stack in mid-2026: the question is no longer whether AI-assisted development is viable. It demonstrably is, for 3 million Codex users alone. The question is how quickly the remaining coordination gaps close, and whether the tool you are using today is positioned to close them.
How to Get Started with the New Features
Update Codex to the latest version through your macOS app store or via the direct download at the Codex product page. All April 16 features ship in this update.
Computer use is disabled by default. Enable it in Codex Settings under the Computer Use section. OpenAI recommends starting with contained, low-risk workflows: simulator-based testing on development builds, GUI-based configuration tasks, and local app testing before using it on machines with sensitive credentials or access to production infrastructure.
The in-app browser is available immediately after updating. For local dev servers, no additional configuration is needed — navigate to localhost in the Codex browser just as you would in Chrome or Safari. GitHub PR integration requires connecting your GitHub account in Codex Settings if you have not already done so.
Memory is enabled by default for accounts that have access to the preview. Navigate to Settings → Memory to review what Codex has stored, edit individual items, or clear the memory entirely. Given that Codex will accumulate codebase-specific context across sessions, reviewing the retention scope before using it on confidential projects is prudent.
Plugins are available in Codex → Plugins. Each new plugin requires separate authorization to access its external service — the connection flow is straightforward and handled in-app. For MCP server integrations, you provide an endpoint URL and authentication credentials for private servers. The full plugin catalog now serves as a reasonable starting point for auditing which of your team’s existing tools can be brought directly into Codex sessions.
Comments · 0
Beta: comments are stored locally on your device and not visible to other readers.
No comments yet. Be the first to share your thoughts.