OpenClaw 2.0 (2026.8.1): What Actually Changed

The OpenClaw project calls v2026.8.1 "OpenClaw 2.0". The useful way to read that label is as a product-wide reset: installation, messaging, memory, skills, models, automation, browser work, native apps, plugins, and security all received coordinated changes.

Sessions become the unit of work

You can search visible conversation text, follow progress through reloads, and run work on paired devices or cloud workers while moving the session workspace with it. Structured questions can return through web cards, native cards, messaging buttons, or plain text, with an explicit skip path.

This makes a long run easier to resume, but it also makes the session a more important security boundary. Check which device or worker owns a workspace before sending a file or granting a tool.

Memory and skills need a review loop

Active Memory can recall bounded personal conversation context, grounded dreaming can promote provenance-qualified material, and automatic self-learning can propose reusable skills. The settings keep opt-out and propose modes, so you can start with review rather than silent mutation.

Keep user-authored instructions separate from generated suggestions. A skill that looks helpful in one workspace may be wrong for another.

Updates are more deliberate

The release documents provider package recovery, model allowlists, session reset behavior, and a set of upcoming SDK migrations. Treat those notes as a migration checklist. Back up state, run Doctor, and verify the Gateway before reconnecting channels.

openclaw doctor
openclaw gateway status

For a line-by-line list, use the official release index. The security documentation remains the right place to decide whether your Gateway boundary matches the people and data around it.