OpenClaw 2026.4.5 adds agent media generation and expands provider support
OpenClaw 2026.4.5 ships video and music generation for agents, expands provider support, and overhauls its memory system.
OpenClaw 2026.4.5 is out. It’s a wide release covering new agent capabilities, provider additions, memory system changes, and a long list of fixes across channels and platforms.
Agents can now generate video and music directly, with results returned inline in the reply. Video support covers xAI, Alibaba Wan, and Runway. Music ships with Google Lyria and MiniMax.
ComfyUI gets its own bundled workflow plugin covering all three media types, for both local and Comfy Cloud setups.
What’s new in agents:
- Video generation: Built-in
video_generatetool with xAI, Alibaba Model Studio Wan, and Runway providers - Music generation: Built-in
music_generatetool with Google Lyria and MiniMax bundled, plus ComfyUI workflow support - Structured progress: Experimental plan and execution item events so compatible UIs can show step-by-step progress during long runs
- Claude CLI bridge: OpenClaw tools now exposed to background Claude CLI runs through a loopback MCP bridge
Provider support got a meaningful expansion this release. New bundled providers include StepFun, Qwen, and Fireworks AI for chat, alongside MiniMax TTS, Ollama Web Search, and MiniMax Search for speech and search workflows.
Amazon Bedrock also gets broader coverage, with Mantle support and inference-profile discovery added. Bedrock-hosted Claude, GPT-OSS, Qwen, Kimi, and GLM routes now require less manual setup.
Other provider additions:
- Embeddings: Amazon Bedrock embeddings for Titan, Cohere, Nova, and TwelveLabs models
- Google Gemini caching: Model-level cache retention support for direct Gemini system prompts on AI Studio
- OpenAI forward-compat:
openai-codex/gpt-5.4-miniadded, plus opt-in GPT personality and GPT-5 prompt contributions - Prompt caching improvements: Deterministic MCP tool ordering, normalized system-prompt fingerprints, and new
openclaw status --verbosecache diagnostics
The memory system gets the most interesting update in this release. OpenClaw calls its memory consolidation process “dreaming,” and it now runs in three phases borrowed from sleep cycle terminology: light, deep, and REM. Each phase handles a different stage of promoting short-term context into durable memory, with REM focused on what the project calls “lasting truths.”
It’s an unusual framing for agentic memory.
Memory changes:
- Dreaming phases: Light, deep, and REM now run independently with recovery behavior
- Aging controls:
recencyHalfLifeDaysandmaxAgeDaysconfigurable per setup - REM preview tooling:
openclaw memory rem-harnessandpromote-explaincommands added - Dream Diary: New UI surface in the Dreams panel
The control UI now ships localized support for 12 languages including Simplified Chinese, Traditional Chinese, German, Spanish, Japanese, and French.
There’s one breaking change. Several legacy config aliases have been removed, including talk.voiceId, talk.apiKey, and channel allow toggles. Existing configs keep working at load time, and openclaw doctor --fix handles migration.
The release also includes a large batch of security fixes across device pairing, gateway auth, plugin routes, and the sandbox, plus channel-specific fixes for Telegram, Discord, Matrix, Slack, WhatsApp, and MS Teams.
Source OpenClaw/Github
Get the core business tech news delivered straight to your inbox. We track AI, automation, SaaS, and cybersecurity so you don't have to.
Just read what you want, and be done with it.
