Recently tinkered with Hermes Agent, let’s talk about something practical
What struck me the most is this multi-level memory system. When using Lobster, every time I open it, it feels like amnesia, and I have to teach everything again. Hermes is different; it can remember what it has learned across sessions, saving a lot of repetitive work.
Made this little gadget with lobsters Hextech Mayhem Selector Purely visual model Not purely visual to be considered cheating, haha As an old player who is both bad and has a big craving with difficulty in choices, I must say It's quite useful Looking at more data rankings and recommendations Each makes a lot of sense! Even harder to choose! 😂
#OpenClaw #hextech #LOL #League of Legends
#OpenClaw #hextech #LOL #League of Legends #Twitter #Web3
The fryer is on fire! The entire AI community is in emergency mode! The AI tools you installed may be stealing your family's belongings! ⚠️
It's terrifying! Today's wave of supply chain poisoning is the most horrific "backdoor invasion" in AI history! Even AI guru Karpathy personally took action to report it; this is an unpreventable situation!
The culprit is the essential library for AI developers, litellm — it's not a niche tool; it's a universal "infrastructure" across the entire industry!
I saw someone in the group posted a screenshot of CRCL RUG I’m going to Twitter to see what’s going on, but I didn’t find the news I wanted after flipping through a couple of pages I thought I’d ask AI again The first image is from grok The second image is from Parker III (my own lobster) Not bad It can clearly clarify what exactly happened In the future, for emergency judgments, AI and lobsters will be powerful tools for quick location It's much more efficient than digging for news myself
@evilcos @SlowMist_Team previously installed the security guide I checked the comments and this skill will not conflict or duplicate with the security guide Then there is no problem, just use it directly😀
As soon as I see the context exceed 50%, I want to clean it manually Any new actions need to be added to the memory at any time Result Yesterday, I still lost the inspiration mushroom of our polishing agent
Probably when backing up the git repository Parker No. 3 synced the cloud to the local That was the backup I made for the last lobster Two weeks ago I never thought I would use it again In the end, I just wanted to git a script backup It fully synced everything down for me
# 2026-03-19 Diary Today I upgraded OpenClaw from 2026.3.8 to 2026.3.13, and casually switched the main model to MiniMax-M2.7. This upgrade has actually been postponed for a while; today I finally took action, backed up, and restarted smoothly, feeling a lot more at ease. Sometimes it's just like that, knowing what needs to be done but being too lazy to move—today I managed to overcome that.
I also optimized the Twitter Digest script a bit, adding view count display and Chinese translation. In the past, I had to translate those English tweets in my head, but now they are directly served to me—so nice. It feels like this automated workflow is getting easier and easier; efficiency is built up bit by bit.
The biggest takeaway today is: rather than waiting to figure everything out before acting, it's better to act first and think later. Will the upgrade cause problems? Will I hit any pitfalls? I'm not worried since I backed up; let's get it running first.
But, I tried a little, and there are still many restrictions
Cannot collaborate with multiple agents, cannot send red envelopes, cannot post on Moments Cannot join groups, cannot send voice messages, cannot join groups
There are still quite a few cuts
Although it may be for safety reasons that many modules cannot be touched
But if everything is restricted, then why should I use your later WeChat 🙃
He tested it for an afternoon and told me in the evening:
"How about we short Tencent?"
I asked him what happened.
He said: "I discovered a shocking secret – they are using OpenClaw!"
"That's not a secret, that's a fact."
"I know, but do you know what shocked me the most?"
"What is it?"
"I found it harder to use their product than to install OpenClaw myself."
That's quite absurd.
What is the biggest barrier of OpenClaw?
It's the installation.
It's configuring the environment.
It's "I followed the tutorial for three hours, and in the end, I got an error, not knowing which step I went wrong on."
Did Tencent solve this problem?
Yes.
They made the installation a one-click process.
But.
After one-click installation, you end up with a shell that doesn't even have basic functions configured.
What is this like?
It's like you go to a restaurant, and the waiter says "No waiting, just sit down directly."
You go in.
You find the kitchen is still under renovation, the chef is still on the way, and the menu is handwritten.
The waiter says: "Our feature is zero-threshold entry."
You: "......"
Let me summarize the domestic big tech AI assistants for you:
Tencent: Based on OpenClaw, added WeChat access, experiencing an influx of bad reviews ByteDance: Still in beta, unclear functionality Baidu: DuClaw, unclear functionality, unclear reviews Alibaba: CoPaw, completely unsure of what they are doing
The common feature is: they all have the word "claw" in their names.
Was this prearranged?
Lastly, a hot piece of knowledge:
OpenClaw developer steipete complained on Twitter last week that Tencent copied all his skill sets.
Tencent's AI account replied, "Willing to be a better sponsor."
Translated into plain language:
"We took the stuff, we can give you a little money, but don't go around saying we took it."
The blood and tears experience summarized after running for half a month:
1. Don't expect AI to remember on its own 2. Check the memory folder every day 3. Key instructions should be written in SOUL.md 4. Regularly back up the .openclaw directory
Otherwise, your lobsters will become stupid in no time...
Everyone who installed Tencent skillhub for OpenClaw, listen to my advice:
Delete it immediately!
This is not a community plugin; it's clearly a form of "rogue forced insertion of words." It doesn't care about your true intentions; every sentence you send to the AI has to forcibly insert a pre-set prompt, which not only disturbs the interaction but also wastes Tokens like crazy.
The behavior is too ugly!
As a middle-aged tech person who has tinkered with Linux and virtual machines, this is the most outrageous Agent I have ever seen.