OpenClaw Upgrade: Video, Dreaming Memory, 12 New Languages

The open-source personal AI agent just became a serious content and automation platform.

OpenClaw, the open-source personal AI assistant you run on your own hardware, has released version 2026.4.5 — and it’s one of the most substantial updates the project has shipped. For anyone building automations, running content workflows, or simply trying to get a capable AI agent without vendor lock-in, there’s a lot here worth understanding.

Video and Music Generation, Built In

The headline addition is native video generation. Until now, getting an AI agent to produce video meant stitching together separate tools and manual handoffs. OpenClaw 2026.4.5 bundles three video providers directly into the platform: xAI’s Grok video generation, Alibaba’s Wan model, and Runway. Users can trigger video creation inside a conversation without leaving the agent environment.

Music generation follows the same pattern, with Google’s Lyria and MiniMax available as bundled providers. For content agencies or anyone running production pipelines, the practical implication is significant: a single OpenClaw agent can now take a brief, generate images, produce video, and create audio — without exporting between tools.

A Memory System That Learns While You Sleep

The dreaming memory system — marked experimental but now shipping with dedicated UI — is the most conceptually interesting part of this release. It runs in the background across three phases: Light, Deep, and REM.

Light phase ingests recent activity and stages candidate memories. Deep phase makes the actual promotion decisions, writing confirmed facts to long-term memory. REM extracts patterns and reflective signals — surfacing themes across what the agent has experienced, without directly writing to long-term memory.

The practical result is an agent that gradually builds a richer picture of how you work: your clients, your preferences, your recurring workflows. A new Dream Diary surface makes this visible and human-readable, and a /dreaming command lets you inspect or control the process. The system is opt-in and off by default.

Prompt Caching, Fixed Properly

Prompt caching in AI systems is one of those features that sounds simple but breaks in subtle ways. This release addresses several root causes that were causing cache misses: duplicate tool inventories in agent system prompts, unstable system prompt fingerprints, and ordering inconsistencies in MCP tools. The fix means follow-up turns are now far more likely to hit the cache, reducing both latency and cost for anyone running OpenClaw at volume.

A new verbose diagnostic mode (openclaw status --verbose) makes it straightforward to see exactly when and why the cache is or isn’t being used.

Structured Task Progress

For long-running agent tasks, the update adds structured progress events — agents can now emit step-by-step status updates that compatible interfaces can display as a real progress view, rather than just a spinning indicator. This matters most for operators running multi-step workflows: product research, listing generation, or anything where knowing which step the agent is on has practical value.

Provider Changes: Anthropic CLI Removed from New Onboarding

One noteworthy change: the Claude CLI backend has been removed from new OpenClaw onboarding. Existing setups that use it will continue to work, and openclaw doctor will help repair or clean up stale configurations. New users setting up for the first time will be directed toward other provider options. The platform remains model-agnostic and works with Claude, GPT, Gemini, and others via direct API or OpenRouter.

Twelve New Languages

The Control UI now supports Simplified Chinese, Traditional Chinese, Brazilian Portuguese, German, Spanish, Japanese, Korean, French, Turkish, Indonesian, Polish, and Ukrainian. For anyone deploying OpenClaw in teams or internationally, the interface is now meaningfully more accessible.

Other Changes Worth Noting

ClawHub — the community library of pre-built skills and workflows — is now searchable directly inside the Skills panel. iOS users can receive push notifications when an agent wants to run a command requiring approval, and approve or deny it from their phone. Amazon Bedrock support has been extended to include Mantle, with automatic IAM authentication removing the need for manual token exports.

OpenClaw is an open-source project. Full release notes for version 2026.4.5 are published at github.com/openclaw/openclaw.