Hermes is easier to use than crayfish. The more you use it, the smarter you become, and you also save Token.

Why did Hermes impress me? Token compression and memory growth. Just these two points I decided to unload the crayfish! Memory and Token Crayfish are not doing a good job in this area. As the conversation grows, the context expands. You can feel that the slower the conversation is, the more expensive the conversation is. In the end, you either manually clear the context or discount the acceptance effect. Hermes is not the same idea at all. It divides memory into four layers, and each layer has different loading timing and Token strategy: Prompt Memory is the resident layer. MEMORY.md USER.md directly injects system prompts, but there is a hard limit of 3,575 characters in total. It is not set casually, but deliberately forces the Agent to refine rather than build up. There is also a counter-intuitive design: memories changed during a conversation don't take effect until the next conversation to prevent temporary information from contaminating the context. FTS5 Session Search solves "how to use history." Save SQLite for each conversation and search for index in full-text. The Agent searches when needed, and the results are injected into the LLM summary without filling the entire history, only the content related to the current task is taken. The Skills layer is procedural memory. The innovation point lies in progressive loading: only names and summaries are loaded by default, and the complete content is adjusted on demand. This means that the Token cost for 200 skills and 40 skills is almost the same. At the top is Honcho, passive user modeling. Agents automatically track your communication style, preferences, and domain knowledge during interactions, and continue to evolve across 12 identity layers through dialect modeling. You don't have to write anything, it's watching you. The effect of the four layers added together: The more the Agent is used, the smarter the Token bill will not rise. ability is relatively Directly compare the three core differences: (1) Closed-loop learning The crayfish Skill is created manually by you, and it won't learn anything new by itself. Hermes is different. When a task meets the trigger conditions (more than 5 tool calls, error recovery, user corrections), the Agent will automatically create a Skill file and record,"This road is clear, there is no need to explore again next time." These skills will also improve themselves in use, updating them with patch rather than rewrite, and only changing the parts that have changed. This is not storage, but accumulation of experience. (2) Token compression The compression of the crayfish is rough, either thrown or stuffed completely. Hermes uses an auxiliary model to trigger compression when the context approaches the upper limit, extract key content to within the 3,575-character limit, and summarize it in intermediate rounds rather than discard it. The key is that the Lineage chain is preserved, and after compression, it can be traced back to the original conversation, rather than irreversible loss of information. (3) Gateway integration Crayfish's Gateway only handles message delivery, and Skill creation, memory writing, and scheduled tasks follow independent mechanisms. Hermes 'Gateway is part of the learning closed-loop: incoming messages can trigger Skill creation, and the output of scheduled tasks is routed back through the same layer. Continue in the conversation terminal where Telegram started, because the Session Binding ID is not bound to the platform. Summary and tutorial I am 100% sure that Hermes is not the optimal solution, but he is the most suitable Agent to solve my problem at this stage! So what is a good Agent? It is a personal assistant who can seamlessly cooperate with you to solve your needs. It is not a hot spot, not traffic, and not a follower of others 'opinions. Whether the Agent is good or not, you must understand it rather than others making it clear! The installation process is very simple. I directly asked Claude code to install it for me. The steps are: python #Install Hermes for me and migrate all the configurations in openclaw, including but not limited to Telegram configurations, model API keys, skills, memory systems, etc. Of course, if you haven't used Claude code yet and want to follow the popularity and install it, then the steps are: python curl -fsSL https://raw.githubusercontent.com/NousResearch/hermes-agent/main/scripts/install.sh | bash After installation, you can use the Hermes setup command to configure it.