Name: OpenClaw 3.11 Update is INSANE! (HUGE FREE RELEASES)
Uploaded: 2026-03-12
Duration: 542 s
Description: Update OpenClaw to 3.11 immediately for the browser-hijack proxy fix, then experiment with the new Gemini multimodal memory and the free Hunter/Healer Alpha stealth models if you want to push context beyond 1M tokens at no cost.

Click to load the YouTube player

Key takeaways

Run `openclaw update` from a terminal in 3.11, not via the bot, so breaking-change warnings (Telegram policy, etc.) surface instead of being silently applied.
Wire up `gemini-embedding-preview-2` with a paid Gemini API key before relying on multimodal memory — image and audio indexing are not on by default.
Index your trade-setup screenshots and green-arrow captures now so semantic queries like "when did my trades work?" return chart matches instead of text guesses.
Try `openrouter/hunter-alpha` and `openrouter/healer-alpha` in OpenClaw within the ~1-week free window if you need sub-1M remote context at zero cost; treat output as agentic-grade, not SOTA.
Skip local Llama routes only if you already run paid frontier models — the Llama first-class setup in 3.11 mainly helps LM Studio users on Qwen-class weights.
Re-test Discord thread management after upgrading; 3.11 ships a partial cleanup, not a full rewrite, so heavy-thread setups may still need manual pruning.

Video summary

Companion notes

OpenClaw 3.11 patches a browser-hijack proxy exploit, adds Gemini multimodal memory embeddings, and ships two free 1M-context stealth models called Hunter Alpha and Healer Alpha on OpenRouter.

## Security: Browser Hijack Fix Version 3.11 closes a way for attackers to get "unauthorized access in certain proxy setups" and gain full control of an OpenClaw instance. The hosts call this the single reason to upgrade tonight, since a hijacked bot exposes the host machine. Update from a terminal with openclaw update rather than asking the bot to self-upgrade, because manual runs surface breaking-change notifications that automated upgrades hide — their own Telegram policy broke during a bot-driven upgrade.

## Memory: Multimodal Embeddings via Gemini The memory system now embeds images and audio into the vector store using gemini-embedding-preview-2. A screenshot of a Pepsi can, a trade-setup chart, or a long voice conversation can be queried semantically — ask "when did my trades work?" and the bot returns screenshots with the corresponding green-arrow setups. It is not turnkey: you must be on a paid Gemini plan, add a Gemini API key in settings, and trigger indexing. The hosts see this as the most useful upgrade for traders and musicians storing recurring visual or audio references.

## Local Models: First-Class Llama Llama.cpp setups get a cleaner integration, so anyone running models through LM Studio (Qwen-class) has an easier on-ramp. The hosts still default to Miniax and Claude Opus for their own bots.

## Free Stealth Models on OpenRouter Two new models — openrouter/hunter-alpha and openrouter/healer-alpha — are pre-wired into OpenClaw. Each offers a 1M-token context for free, remote, no local hardware. Early tests are "not super good"; the hosts believe the models are stealthily optimized for OpenClaw agentic workflows. Expect them to remain free for roughly one week before an official reveal.

## Minor Fixes iOS UI polish, ACP developer improvements, and Discord thread-management cleanup.

Watch on YouTube

Prefer the native player? Open it on YouTube: https://www.youtube.com/watch?v=goy7Ob0I5D8