OpenClaw 2.0 adds team sessions and more memory for agents

In the browser, OpenClaw 2.0 no longer begins with a configuration marathon, but directly with a conversation. Users can complete setup there, resume tasks and follow the AI agent’s work live. The new version has arrived after a nearly seven-week release pause — and it bundles the work of 933 developers and nearly 17,000 pull requests.

The most noticeable step takes OpenClaw beyond the individual user. Shared Cloud Sessions allow multiple users to work in the same ongoing session or hand off tasks without losing the existing context. The development team calls this feature “Multiplayer” and says it is already using it itself for project work. OpenClaw now also recognizes existing ChatGPT or Claude subscriptions, API keys and local models during fresh installations.

The agent’s memory is expanding as well. In personal installations, Active Memory can retrieve selected information from earlier private conversations with the same agent. Grounded Dreaming, active by default, is intended to condense relevant content in the background into longer-term memories. A self-learning system records reusable insights and can automatically deploy verified capabilities or those managed by Skill Workshop.

For automation, this means fewer repeated approvals: recurring actions are supposed to continue after one-time approval, and a plug-in for incoming emails can start agent work. Connected Macs can already be controlled through desktop interfaces. On Windows and Linux, by contrast, this computer control is still experimental and must be explicitly enabled. Conversations can be searched, branched or rewound; this does not undo file changes or other tool actions that have already been executed.

So what does this change in practice? Individual users can set up OpenClaw more easily and automate longer-running tasks more extensively. Teams can hand off work within a session without rebuilding the context. At the same time, the agent’s reach remains a security problem: prompt injection — manipulated instructions in untrusted content — can become dangerous when powerful tools are accessible. OpenClaw does introduce new safeguards for credentials, plug-ins and shared sessions; sandboxing, however, is disabled by default. Compared with offerings from OpenAI, Google, Anthropic and Microsoft, the practical difference remains above all the combination of open source, self-hosting and freedom of model choice.