OpenClaw Update 4.12 is IMPROVING!

OpenClaw 4.12 lands active memory, Codex routing, and three security patches in a single day after yesterday's release.

Active Memory Plugin: Executive Assistant Behavior

The 4.12 headline is the active memory plugin, modeled on Honcho's memory layer. A dedicated sub-agent fires automatically before every reply, searches stored context, and injects relevant bits into the prompt — no more telling your agent "remember this" first. You can tune it with recent message or full context modes. The creator's example: if you recently switched to keto, the plugin should combine the diet switch with past lunch orders and recommend a keto meal without being asked. That's the executive-assistant bar — and the creator points out real human EAs cost $2,000–$5,000/month.

How To Audit What Got Pulled In

If you don't trust what the memory layer injected, type /verbose to see exactly which context the agent loaded. The creator calls this a habit you "should" keep, framed as "if you still have trust issues with your OpenClaw, which you should, by the way."

Two New Model Providers

First, Codex/ChatGPT models now route through Codex's managed API with native threads — a path completely separate from the standard OpenAI route. The framing is direct: "Anthropic, you disrespected us in the past" (Claude Code disabled easy off-management), so OpenClaw is leaning into the GPT ecosystem. Second, LM Studio local inference is supported, with Gemma 4, Qwen 3.9, and a Qwen 3.5 Opus-distilled reasoning model as recommended picks for free local runs.

Plugin Loading & CLI Cleanup

Plugins now only load what their manifest declares — a fix the creator says "should have been done in the get-go." New CLI: openclaw policy with preset and set sub-commands for managing tool approval policies from the terminal.

Security Patches

Three critical patches shipped. BusyBox and ToyBox were removed from the trusted safe-bins list, closing a shell-escape vector. Media-fallback is also fixed, which had been dropping messages after the creator disconnected WhatsApp from the Gambit agent.

Why The Creator Hasn't Migrated Back

The team moved to Hermes for self-evolving skills and verbose behavior. Active memory plus dreaming gives them "sufficient reason to just stay… keep warm" on OpenClaw, but the verdict is: warming, not switching.