OpenClaw 4.12 update is actually incredible
TL;DR
OpenClaw 4.12’s biggest win is memory that actually compounds — Alex says the new 3:00 a.m. “dreaming” pass plus active memory checks and his Obsidian daily-log setup made his OpenClaw “forget nothing,” because it now reviews the day and promotes useful details into permanent markdown memory.
Skip third-party memory plugins and keep everything local — He argues tools like Mem0 undermine OpenClaw’s whole value prop because memories move into external databases/cloud systems instead of staying in local markdown files your agent can fully own and customize.
The smartest OpenClaw setup is multi-agent, not tool tribalism — Rather than picking sides in the OpenClaw vs. Hermes fight, Alex says the real move is using Opus-powered OpenClaw as orchestrator and delegating coding to Hermes or the new native Codex plugin to save money and recover when one agent breaks.
MIT was his proof that AI power users are way earlier than they think — After speaking at MIT and talking with investors, billionaires, and former Obama-era U.S. CTO Megan Smith, his takeaway was that almost nobody there was using OpenClaw deeply, which he says means regular viewers are ahead of many elite operators.
He thinks AI Twitter is flooded with illegal-ish astroturfing — In a long rant sparked by Lovable Payments quote tweets, he says the “Stripe CEO right now” reaction-GIF posts are obvious paid promotion without disclosure and a big reason you shouldn’t trust hot takes like “this tool is dead.”
He’s unusually blunt about model tradeoffs right now — Alex says Opus 4.6 has gotten noticeably worse over the last few weeks, but he still prefers it for agentic coding because GPT-5/ChatGPT remains too reluctant to finish tasks, describing the gap as Claude relentlessly crawling to the finish line while ChatGPT gives up early.
The Breakdown
The 4.12 update starts with “dreaming” and a very Alex explanation
Alex opens by calling OpenClaw 4.12 “absolutely massive” and immediately spotlights the new dreaming feature: every morning at 3:00 a.m., the agent reviews the day and decides what deserves permanent memory. He leans hard into the metaphor — your OpenClaw now “dreams” like a person does — and says paired with his Obsidian daily-log system, it makes memory “flawless.”
Why he refuses to use third-party memory plugins
From there he gets practical: don’t bolt on outside memory systems like Mem0 if you care about OpenClaw’s local-first setup. His point is less anti-plugin than anti-outsourcing — if a memory tool stores your life in someone else’s database, you’ve broken the point of having a private markdown-based agent in the first place.
Active memory and the case for custom-built systems
The next major feature is active memory: before every reply, OpenClaw now checks its logs, dreams, and permanent memory to see what context matters. Alex says this may cost a few more tokens, but the tradeoff is worth it because the assistant is much less likely to forget what you’ve already discussed.
OpenClaw vs. Hermes is the wrong argument
A big chunk of the stream turns into a fight against AI Twitter tribalism. He says Hermes is better than OpenClaw in some ways and OpenClaw is better in others, and the winning move is using both together — not declaring one “dead” every time a new feature drops.
The astroturfing rant: why he says Twitter is basically unusable
That spills into his most heated segment: a takedown of what he sees as obvious paid quote-tweet campaigns around Lovable Payments. The repeated “Stripe CEO right now” reaction-GIF posts are, to him, textbook undisclosed shilling, and he says this is exactly why AI Twitter news and benchmark chatter can’t be trusted at face value.
Models, motivation, and why Claude still wins for agent work
Asked about model quality, he says something he used to dismiss: Opus 4.6 really does feel dumber now. But even then, he still picks Claude/OpenClaw for coding because GPT-5 and ChatGPT feel brilliant-yet-unmotivated — his memorable analogy is ChatGPT as a stoner who quits after stubbing its toe, versus Claude as a meth-fueled runner who keeps crawling toward the finish line.
Codex plugin, ChatGPT memory import, and the beet-juice detour
Back on 4.12, he highlights the native Codex plugin, which lets OpenClaw spin up Codex agents directly for cheaper coding execution while Opus orchestrates. He also calls the new ChatGPT memory import a sleeper hit: years of ChatGPT context can now be pulled into OpenClaw, then filtered through dreaming into permanent memory — though he briefly interrupts all this with a very live-stream story about drinking beet juice, thinking he was dying after seeing red urine, and then realizing ChatGPT had the answer.
MIT, Henry, and his bigger thesis about being early
Late in the stream, Alex pivots to his MIT keynote on AI agents in distribution and what he learned talking to investors and high-status tech people. His conclusion is pure FinnFam energy: if you’re seriously using OpenClaw, Claude Code, and local workflows, you are probably ahead of a shocking number of billionaires and MIT insiders — and that’s the reason to keep going, keep building, and keep stacking leverage while he ramps Henry Intelligent Machines in parallel.