SectorHermes Is it working
Hermes andOpenClaw What difference does it make
Give up the lobster turn, Hermes, wait...
It's supposed to be the most subjects and questions we've seen in the last few days, but most of the questions lie behind us and we don't know what we really want
Lobsters and Hermes aren't on the same track
If it's a choice between lobster and Hermes, it's like a question of whether to buy Iphone or Ipad. They're all common Agent The system, that is uncontroversial. It's based on big models, Loop loops, Harness operating frames, and so on, but it's completely different
Through two weeks of depth, today I try to make their differences clear. There's a proposal behind the article
Saved:
Lobsters are good at entry and order management. It's like a company's professional front and corporate system, who can come in, how the mission is divided and who can do what cannot。
Hermes is good at settling in experience
One sentence: Lobster makes it possible for your AI to be found and to work in a regular manner. Hermes makes your AI increasingly skilled。
It is very simple to say, but the gap between the two is much greater than we thought. Let me summarize the most important tripartite aspects:
I. SKILLS: DISTINCTION BETWEEN SOP TREASURY AND WORK NOTES
Lobster skills are human machine rules, installation of Skill, definition of functional boundaries, engagement of input output and implementation logic, and local tiered governance of the system. Anyway, it's not bad. It's all in your head. That's why many people use lobster badly。
Hermes ' skills, in addition to being manageable, are most critical to learn on their own, to summarize themselves, to automatically preserve successful steps as available skills, and to record transactional processes for specific services。
As a matter of fact, lobster is a standard manual for the team, and Hermes is a personal experience note
Memory: the difference between the solidification of memory and the brain of the search engine
The memory of the lobster is based on the optimization of local md documents, where the clarity of the combing is unclear, and how the reference directly affects the memory quality of the lobster. Soul.md defines character, user.md records preferences, memory.md pre-selections, logs filed by date, etc. I bet the 801 TP3T lobster player didn't design to optimize his memory system, even more catastrophic if it was natural。
Hermes uses three layers of memory: conversational memory, lasting memory, skills memory。
It's not a memory, it's a work-by-work experience, it's for next time. The complete memory is closed through four steps:
Work, run the whole process of your mission。
RECORDING, WHAT YOU SAY, WHAT YOU DO, HOW YOU ADJUST IN THE MIDDLE, WHAT THE END RESULT IS, ETC. THE INDEXING USING FTS5 IS SUSTAINABLE AND IS SEARCHABLE IN ITS ENTIRETY. AND IT'S NOT STACKED TOGETHER, IT'S STORED IN THREE LAYERS:
- –Memory of the session: context of the current conversation, ensuring consistency of answers - lasting memory: long-term user preferences and patterns of behaviour, next time you are recognized - skills memory: path to the operation of a successful mission, automatically sealed as reusable
3. Sedimentation techniques, which automatically capture “how to make this happen” upon completion of complex tasks. Next time a similar mission is conducted, there is no need to check and directly call on existing skills。
The user modeling, analysing your long-term behaviour and preferences, slowly learning what you like, what you're taboo about, and the answers increasingly fit you。
So the biggest difference between Hermes and lobster memories is how much remains and how much can be recovered next time。
Security: Distinction between pre-emptive rules and interception in the event
It's basically a recapitulation of what I used to do when I was safe. Lobsters are pre-assembled, pre-empted rules, white lists, etc., but it is difficult when they land, because no one can imagine。
We have dealt with problems in the past, in the first, in the middle, and in the second. Hemes uses a programme that is managed in a way that assumes, in a zero-confidence way, that the environment is in question, so it uses a deep-default horizontal separation model, high-risk manual clearances, automatic scanning in the context, and so on, and how to intercept when the main attack is a dangerous operation。
- Shit, I've got a feeling of going back to the old writing technology
Summarizing, from a security point of view, the heavy management of lobsters, Hermes the heavy defence。
So, what about the two of them
I insist on the same point of view. Don't listen to what others say. Let's figure out what you want to use
If you're alone and expect Agent to understand you more and more, choose Hermes
If you have a complex multiAgent access, you need teamwork, process standardization, lobster selection
Lobster can help you access complex environments, Hermes can help you sink complex experiences. They're not two or one, they're one that suits you better. Harnesses has been very hot lately, and lobsters and Hermes are not the end