OpenAI Dots Explained: The Always-On Agent With Memory You Can't Control

OpenAI used its DevDay keynote on September 29, 2026 to make its biggest bet yet on agents that keep working when you are not looking. The product is called Dots: always-on, personal agents powered by the GPT-6 Astra model, each with its own cloud computer and browser, pursuing goals in the background with minimal oversight.

If you run scheduled or recurring AI workflows, Dots matters to you even if you never touch one. This launch is OpenAI confirming what the automation crowd has said for years: an agent that wakes up blank every morning is a toy. Memory is the product. But OpenAI's implementation of memory for Dots has a serious limitation buried in the fine print, and that limitation is exactly where an independent memory layer like Vilix AI fits.

What Dots actually are

A dot is a persistent agent you set up inside ChatGPT. Give it a name and a goal, and it keeps working on that goal continuously. Unlike a chat that waits for a prompt, a dot performs what OpenAI calls "proactive research": it reads connected apps with read-only tools, forms notes about what it finds, and pursues ongoing responsibilities like monitoring customer feedback or tracking a recurring pipeline.

Each dot gets its own cloud computer and its own browser. It connects to more than 4,000 applications through OpenAI's plugin ecosystem. You reach it through ChatGPT, Slack, and Microsoft Teams, with text message support on the way. The first dot is included with ChatGPT Pro and Business Premium plans, and OpenAI is positioning specialist dots with their own identities, credentials, and tools as the next step.

Where Dots gets memory right

The memory design is the most interesting part of the launch. A dot draws on two sources: ChatGPT's existing memory of your preferences, and its own notes about your ongoing responsibilities and working habits. Context carries across ChatGPT, Slack, and Teams, so a preference stated in a chat can inform what the dot does overnight. This is precisely how a serious agent should behave: the job survives the night because the memory does. It also proves what scheduled-agent operators have said for years: the model alone was never enough. The winning product is the model plus persistent memory plus tools, waking up with context already loaded.

The catch: you cannot manage what a dot remembers

In its Dots privacy, security, and safety FAQ, OpenAI states that users currently cannot view, delete, or directly modify individual dot memories, including details that entered the dot's context from plugins. The only way to delete a dot's context is to delete the dot itself, which wipes its conversations, saved memories, and scheduled tasks together.

Disconnecting an app does not help either. OpenAI's documentation says disconnecting a service stops new access, but it does not delete information the dot already built into its context. A dot that read a shared inbox or a CRM overnight keeps what it learned even after you cut the connection. When you want one customer removed from what your dot knows, you cannot remove one customer. You can only delete the whole agent and start over.

That is an all-or-nothing memory design. No per-memory delete, no editing a wrong note, no auditing which detail came from which source. For a personal assistant this might be acceptable. For an agent that reads business systems on your behalf, it is a real operational problem: wrong memories cannot be corrected, stale ones cannot be retired, and weeks of accumulated context cannot be exported or moved.

The second problem: the memory never leaves OpenAI

The dot's memory works across ChatGPT, Slack, and Teams. That is the entire list. If your automation stack includes n8n workflows, a scheduled reporting job, a separate coding agent, or any tool outside OpenAI's walls, the dot's memory never leaves OpenAI's garden. The context your dot spent weeks accumulating is not portable, not queryable from your other agents, and not inspectable by you. It is the classic platform lock-in pattern moved one layer up: previously your documents were locked in a vendor; now your agent's accumulated judgment is.

The missing layer: portable memory any agent can read

Scheduled-agent operators solve this with a different architecture: one memory layer that every agent reads, regardless of which company made the agent. Vilix AI is built for exactly that.

Vilix AI is a cloud-hosted memory layer with zero infrastructure to run. Connect any MCP-compatible AI client to your Vilix AI account (Claude, Codex, Cursor, OpenClaw, Hermes, GitHub Copilot, Windsurf, Lovable, Muse, and any other MCP-compatible tool), and every agent reads and writes the same shared memory. Plan in one tool, build in another: the context, rules, and tasks come with you. Retrieval is semantic plus keyword search, so agents find what was meant, not just what was typed.

Critically, the memory management works the way Dots' does not. You can list, update, and delete individual memories from any connected AI or the dashboard at app.vilix.ai. One wrong note gets corrected in one place, and since every agent reads the same memory, every agent sees the correction. Conflicts resolve with last write wins and recency-aware retrieval. You can export all of your memory in a portable format anytime, or wipe the account instantly. No deleting the whole agent to forget one detail.

The model is still the model. A dot running GPT-6 Astra on an OpenAI cloud computer is a capable worker. But memory is the part that compounds: the preferences, the decisions, the accumulated context of months of runs. Tying that value to a single vendor's walls means rebuilding it from scratch the day you add a tool outside those walls. A vendor-neutral memory layer means the context follows the work, not the subscription.

Honest framing

Dots is a genuinely strong product: always-on agents, a cloud computer each, read-only proactive research with custom rules and activity views, on a frontier model. For work that lives entirely inside ChatGPT, Slack, and Teams, a dot is a compelling package.

The tradeoff is control and portability. Today you cannot inspect, edit, or delete individual dot memories. The accumulated context cannot leave OpenAI's ecosystem. And disconnecting a data source does not erase what the agent already learned from it.

Vilix AI's tradeoff is the mirror image. It does not build or run the agent, and it is cloud-hosted only. What it does is keep one portable, fully manageable memory that every agent in your stack reads over MCP: full conversation history, per-memory view and delete, export anytime, a free plan that stays free, and a 7-day Pro trial with no card required.

The agent era needs both layers. Dots proves the always-on agent has arrived. The question is whether its memory belongs to you or to the vendor. For agents that matter, the answer should be you.

Sources: TechCrunch reporting on the Dots launch at DevDay 2026; VentureBeat on Dots and ChatGPT Space; OpenAI's Dots privacy, security, and safety FAQ (help.openai.com); Unite.ai coverage of plan availability and memory reset behavior.