Continuing work assistants

These are assistants with an ongoing job. In the documentation, the four closest to "a bot that learns its job and keeps working" are Grok Bot, OpenAI dots, Meta Muse and Claude Duet. OpenClaw and Perplexity also belong here. Each entry separates what the documentation says from our assessment.

Product	Organizing idea	What carries forward	How it works	What sets it apart
Grok Bot	A roster of named bots with different jobs	Each bot's conversations, learned role context, preferences and summaries; shared files and browser sessions	A persistent cloud computer with a browser, filesystem, terminal, connected apps, parallel bots and handoffs	The most explicitly organized around several continuing specialists
OpenAI dots	An always-on assistant that learns your standards and moves ongoing work forward	Its own retained context, plus memory and recent context shared with ChatGPT	Its own cloud computer and browser, connected apps, background work, proactive research	Strong ChatGPT continuity, but you can't inspect or edit its individual memories
Meta Muse	A personal agent that advances goals and learns about you	Conversations, goals, scheduled tasks, artifacts, and editable memory and personality files	A dedicated cloud Linux environment and browser, connected services, background work	The most explicit file-based memory controls of these products
Claude Duet	A proactive professional assistant built on your existing Claude context	Claude memory, skills, connected apps and one ongoing main thread	A cloud virtual machine and browser, scheduled account checks, background tasks	Aimed at professionals and small businesses, with clearly stated approval boundaries

Sources: docs.x, openai, ai.meta, claude

These are differences between products, not just between the underlying models. For example, Grok's bots share one account-level computer, while OpenAI describes each dot as having its own cloud computer. That affects how bots coordinate, how isolated they are, and how you should organize confidential work. docs.x

Grok Bot

What the documentation says. It describes named bots that keep role-specific context. They work in parallel, message one another, join group conversations and hand tasks to each other directly. They have a persistent browser, filesystem and terminal. Workflows can be learned from a demonstration, saved as skills and rerun on a schedule. docs.x

Memory. Each bot can keep stable preferences, role context and summaries of earlier work. Conversations and learned context stay separate for each bot. Shared files, browser sessions and handoffs move information between bots. Grok's own guidance says consequential decisions should be checked against current sources, not against memory. docs.x

The caveat. All the bots on an account share one cloud computer, including its files, browser sessions and app logins. Separate bot names don't create separate security compartments. Each bot gets its own screen, but the screens are not separate accounts or machines. docs.x

Our assessment. Grok Bot is the closest match to a team of continuing specialists, such as a project coordinator, a supplier researcher and a proposal reviewer. The roles are separate, but they share one working environment. That suits a team working on one business's projects. On its own, it isn't enough to isolate confidential clients.

Platforms and individual bots. A bot template runs inside a platform; it isn't a platform. Compare platforms with platforms, and specialist bots with other specialist bots on the same platform. The template listings show the platform's categories: memories, skills, routines and integrations. x

OpenAI dots

What the documentation says. OpenAI presents dots as always-on agents that learn what matters to you and keep working toward goals. The launch describes GPT-6 Astra, a cloud computer, a browser, connected apps and learning from feedback. openai

Memory. A dot can receive ChatGPT memories and recent conversation context. In turn, dot conversations can add to ChatGPT memory, and this sharing can continue after setup. Turning off ChatGPT memory stops the sharing, but it doesn't erase what the dot already received. OpenAI currently says you can't view, delete or directly change individual dot memories. Deleting the dot deletes its own context. Separately stored files, Codex threads, ChatGPT conversations and memories already shared into ChatGPT each have to be managed separately. help.openai

Research and actions. Dots can read permitted connected sources on their own initiative and save private notes. Those research tools can't send messages, change content through plugins, or control a browser or computer. Any follow-up action has to pass the normal action rules and safety checks. Dots also offer Custom Rules, action review, and an Activity View of ongoing and delegated tasks. Some recurring actions can be approved in advance, within limits you set explicitly. help.openai

Our assessment. "Always on" doesn't mean unrestricted action. Dots are a strong choice when continuity with ChatGPT matters most. But memories you can't inspect are a weak official record for requirements and exceptions. An editable project document does that job better.

Meta Muse

What the documentation says. Muse is a personal agent that can advance longer-term goals, open a browser, fill in forms and keep working after the app closes. It asks for approval before sensitive actions such as sending email or making purchases, and it runs in a dedicated cloud environment. about.fb

Memory. Meta documents separate files that you can edit: Memory.md holds remembered information used for personalization, Soul.md holds core truths, boundaries and personality, and Identity.md holds the agent's identity and presentation. Goals, artifacts, system files and scheduled tasks are shown separately, and you can export chats, files and other agent information. Deleting a message or file doesn't necessarily remove everything Muse learned from it. Meta says its "forget" skill removes information "to the best of its ability." meta

Security and privacy. A separate Sentinel agent checks outbound activity and asks for permission when needed. Meta says Muse doesn't share conversations or VM data with its advertising systems. about.fb The setting that lets your interactions be used to improve Meta's models is on by default. meta

Our assessment. Muse gives one of the clearest documented answers to the question "Does it write things down, and can I see what it thinks it knows?" Turn the model-improvement setting off before you test with proprietary or client material.

Claude Duet

What the documentation says. Duet is aimed at professionals and small-business owners. It inherits Claude memory, skills and connected apps. It sends a morning brief, checks connected accounts every few hours, and works in a cloud VM while your devices are off. It is designed around Claude Opus 5.5. Duet keeps one ongoing main thread and can return work such as documents, dashboards, financial models and prototypes. Duet was announced on October 21 and is a research preview. claude

Approval boundaries. In the preview, Duet asks before it sends messages, makes payments, signs documents or deletes anything. Some actions are prohibited even with approval, including trading securities and moving money between accounts. Access starts with a waitlist for US Pro and Max subscribers. Duet shares usage limits with the rest of Claude, and its background checks use up those limits. You can't view its browser yet. Anthropic warns that it may misread a thread or misjudge urgency. claude

Our assessment. For an existing Claude user, the main advantage is not having to rebuild context, skills and connections. The announcement doesn't explain how Duet picks memories, resolves memories that contradict each other, or versions the exceptions it learns. So don't read "carries over your memory" as a promise of database-like reliability.

Also in this group

OpenClaw

What the documentation says. OpenClaw is an open-source assistant that runs on your own machine and works through messaging channels such as WhatsApp and Telegram. It handles background tasks, reminders and recurring work. openclaw Its durable memory is a set of Markdown files. MEMORY.md holds curated long-term facts, standing decisions and short summaries. Daily notes keep more detail, and useful parts of them can be distilled into long-term memory. The documentation puts it plainly: the model remembers what is saved to disk, and there is no hidden durable state. If the memory file grows past its startup budget, the file stays intact on disk, but the copy loaded into context is cut short. docs.openclaw Its automations can keep scheduled jobs, wake the agent, and deliver results to a channel or webhook. They support intervals, cron expressions, one-time reminders and event triggers. docs.openclaw

Our assessment. OpenClaw is the closest match to a continuing coordinator that you run yourself. Saving information and reliably finding the right piece of it again are still two separate problems.

Perplexity Computer, with Projects, Brain and Automations

What the documentation says. The unit to compare is the combination, not Computer alone. Projects hold continuing work, shared files, instructions, tools and tasks. perplexity Brain studies a project's sessions and files and keeps an evolving memory, which you can view. You can choose whether it updates automatically or manually. perplexity Automations run on a schedule or on qualifying events, use connected files, tools and Projects, and keep context from earlier runs. They replace Scheduled Tasks. perplexity

Our assessment. It fits project coordination well: checking open items against what changed since the last run. Look closely at how Brain's learned context and an automation's run history interact. Neither one should be treated as an approved decision log.