OpenClaw 2026.8.1 turns memory on by default and drops the bundled OpenProse plugin

OpenClaw published version 2026.8.1 on August 31, 2026. For anyone already running it, the change that matters most is not a new feature but a changed default. Three things that used to be switched on by hand now happen on their own: the agent recalls earlier private conversations, consolidates memory in the background, and applies new skills that its own scanner has cleared.

Three defaults changed, and all three touch memory

Personal conversation recall comes first. On personal installs with no direct-message isolation configured, and with Active Memory enabled, the agent now pulls bounded context out of earlier private conversations with the same agent. Groups and channels stay excluded, and an explicit control switches recall off.

The second is what the release notes call grounded dreaming: model-backed memory consolidation that runs in the background, is enabled by default, and promotes provenance-qualified material into long-term memory. It keeps a Dream Diary and ships with an explicit disable control. The third is automatic self-learning, which captures reusable lessons and applies scanner-approved skills without a prompt. Skill changes a person wrote still wait for approval.

• Recall of private conversations is on by default on personal installs, and leaves groups and channels alone.
• Background memory consolidation is on by default and writes a Dream Diary.
• Scanner-approved skills apply themselves, while changes a person authored still wait for approval.
• On macOS, Peekaboo and Allow Computer Control stay on by default, subject to operating system permissions.

Two migrations break setups that worked yesterday

The bundled OpenProse plugin and the /prose command are gone. Source files with the .prose extension are kept, but the configuration that pointed at the plugin is now stale, so the release asks operators to run openclaw doctor --fix and to follow the upstream migration to the Agent Skill format.

The second migration rewrites model references. Shipped codex/* and openai-codex/* routes, provider configuration, stored sessions, and automation routes all move to openai/*, again through openclaw doctor --fix. Conflicts are reported to an operator instead of being resolved silently. Separately, authors of external plugins have a deadline of September 1, 2026. On that date several SDK import paths move into narrower namespaces, though nothing is removed in this release.

A credential the model never sees

The release adds private credential requests. An agent can ask for a secret value through a masked prompt, and that value enters neither the conversation transcript nor the model context. An opt-in proxy limits the destinations a protected secret may be substituted into.

For teams this is backed by a shared credential store in SQLite, where secret values are write-only and protected outbound traffic is bound to declared hosts. Recurring approvals follow the same shape: an operator approves one exact operation, can inspect or revoke it later, and a changed job or operation requires fresh approval.