What Is Manus AI? How the Autonomous Agent Works
Manus is a general AI agent: you give it a goal, and it works through the steps on its own cloud computer. It launched in early 2025 and grew fast. Meta acquired it in December 2025. In 2026 Manus separated from Meta and returned to independent operations. Manus 2.0 shipped on September 28, 2026. Below is what Manus AI is today and how it works. We also look at how it keeps context within and across tasks, which its team has documented in unusual detail.
What is Manus AI, in one sentence?
Manus is an autonomous general-purpose AI agent that plans multi-step tasks and carries them out on its own cloud computer.
In a July 2025 post, Manus said its early-2025 launch "defined the category of General AI Agents." Work in Manus happens in tasks. Each task runs in a virtual computer the agent controls. Manus announced it was joining Meta on December 29, 2025. By then, it said, its agent had processed more than 147 trillion tokens and powered more than 80 million virtual computers.
How does Manus differ from a chat assistant or a coding agent?
| Dimension | Manus | Chat assistant (ChatGPT, Claude chat) | Coding agent (Claude Code, Codex) |
|---|---|---|---|
| Main job | Finish a task end to end: research, documents, apps, slides | Answer and draft in a conversation | Change code in a repository |
| Where it works | Its own cloud computer, or an authorized computer of yours | The chat window, plus tools the vendor adds | Your machine or the agent's cloud workspace |
| Unit of work | A task that can run many steps | A message | A session or a task |
| Context across tasks | Project instructions, files and skills | The vendor's memory feature, such as ChatGPT memory | Instruction files such as CLAUDE.md or AGENTS.md, plus host memory |
| Who controls the model | Manus by default, or your own provider through Manus Flex | The vendor | The vendor, or a model you pick |
Sources: the Manus 2.0 launch post and the Manus Flex post.
How does Manus work?
You describe the outcome, and Manus works out the steps. Manus 2.0 runs on a new agent harness the company calls Cascade. Per the launch post, it starts each project lean. Specialized tools load only when a task needs them. Manus says one tested setup used 23.2% fewer tokens and cost 32% less to run than the previous system. That is a vendor benchmark without a published method.
Around that core, Manus 2.0 adds products for specific jobs. A paid Cloud Computer gives projects persistent hosting. Automations start work on events from connected services, such as an email or a calendar change. Remote Control and Computer Use let Manus work on a computer you authorize. The Manus Studio desktop app puts documents and code in one shared workspace.
For research over long lists, Manus uses Wide Research. Its October 2025 post explains the design. Each item goes to its own sub-agent with "a fresh, empty context window." Ten items get ten sub-agents, and 500 items get 500. A main controller collects the results and writes the report.
Manus Flex, announced on September 29, 2026, lets you plug in your own API key from OpenRouter, Fireworks or Modal. Your provider bills you for model inference. Other services used in a task still consume Manus credits, so Flex lowers the bill only for the model part.
How does Manus handle context inside a long task?
We think Manus has published the most useful public account of how a general agent manages context. Yichao "Peak" Ji wrote Context Engineering for AI Agents (July 18, 2025) after the team rebuilt its agent framework four times. A typical Manus task then took about 50 tool calls. Input outweighed output by about 100 to 1.
The main lessons:
- Protect the KV cache. Ji called cache hit rate the most important metric for a production agent. With Claude Sonnet at the time, cached input cost $0.30 per million tokens versus $3 uncached. Manus keeps the prompt prefix stable and the context append-only.
- Mask tools instead of removing them. Adding or removing tools mid-task breaks the cache and confuses the model. Manus constrains which actions the model can pick at decoding time.
- Use the file system as memory. The post calls files "unlimited in size, persistent by nature." Manus drops a web page from context but keeps its URL, so it can read the page again.
- Recite the plan. Manus rewrites a todo.md file as it works. That keeps the goal near the end of the context, where the model attends most.
- Keep errors in context. A failed action and its error help the model avoid repeating the mistake.
In our view, every lesson except the second applies to any long-running agent. The second assumes you can constrain decoding, as Manus does.
How does Manus remember things across tasks?
Within one task, the state lives in the context window and in files. Across tasks, Manus uses Projects, introduced on December 1, 2025. A project has a master instruction that every new task follows, plus a knowledge base of files such as brand guidelines or documentation. Collaborators you invite inherit the project's context. Changes to files apply only to new tasks.
In May 2026, Manus made projects learn from their tasks. After a task, Manus can review the conversation and propose updates to the project's instructions and skills, with a reason for each change. Nothing changes until you approve it, and approved updates apply only to that project.
We think human approval is right for shared instructions. The cost is that what one project learns stays in that project, inside Manus.
What should you read in the fine print?
The usage and performance figures here are Manus's own. The token and virtual-computer counts come from the Meta announcement. The Cascade savings come from an internal test. We found no independent evaluation of Manus 2.0 in the sources we checked.
The Meta chapter is the bigger caveat. On August 11, 2026, Manus told users it was returning to independent operations and separating from Meta. To meet regulatory requirements in some jurisdictions, it would delete data that certain users generated on or after December 29, 2025. Affected users could back up their data before August 23 and restore it from August 25. On September 1, 2026, Manus said it had returned to self-governance as "an independent agent lab," with the founders in charge. Users who backed up but have not restored can still do so, with no deadline.
Manus said no breach caused the deletion. The cause was a change in corporate ownership, which is a risk for any context a vendor holds.
The bottom line
Manus is worth trying for long, multi-step work outside a code repository. Good fits are research across many items or a small app. Put recurring work in a Project and review the update proposals it makes. Keep your own copy of instructions and knowledge that matter. The 2026 data deletion showed that vendor-held context can disappear for reasons outside your control.
How do you keep agent memory outside one vendor?
Manus manages context well inside a task, and Projects carry instructions across tasks with your approval. Both live in one vendor's system and are scoped to one project. When Manus separated from Meta, some users had to back up and restore their work. And when you switch to a coding agent for the repository side of a project, nothing Manus learned comes along.
memU keeps memory as a shared wiki of readable Markdown files. A scheduled background task slices new session logs from each supported host into jobs. The agent decides what is worth keeping and writes it as memory or skill Markdown. memU's memory service makes no LLM calls: it stores and retrieves what the agent prepared. All hosts share one backend, so Codex and Claude Code read the same memory. You can self-host it on a single device for privacy. memU does not connect to Manus. It covers coding agents such as Claude Code.
To try the hosted version, which is free and syncs across devices, get an API key from memu.so and send your agent one message:
Read https://memu.pro/SKILL.md, follow its instructions to install and configure memU, API Key is
YOUR_MEMU_API_KEY.
For the self-hosted setup and the support list, see the memU GitHub repository.
FAQ
Is Manus owned by Meta?
Not anymore. Meta acquired Manus on December 29, 2025. In August 2026 Manus announced it was separating from Meta, and on September 1, 2026 it said it had returned to independent operations.
What is Manus 2.0?
Manus 2.0 is the version announced on September 28, 2026. Built on a new agent harness called Cascade, it adds a paid Cloud Computer and event-based Automations.
Does Manus have memory?
Manus keeps context within a task through its files and context window. Across tasks, a Project holds a master instruction plus knowledge files. Manus can propose updates to them, which you approve.
Can I use my own model in Manus?
Yes, through Manus Flex. You connect an API key from a supported provider such as OpenRouter. That provider bills you for inference. Other parts of a task still use Manus credits.
What is context engineering in Manus?
It is how Manus decides what goes into the model's context on each step. Its team described four habits: keep the KV cache stable, use files as external memory, recite the plan in a todo.md file, and leave errors in context.