Use /goal
Got a question about how you use Engram. Do you create a Python service with HTTP API ? Do you store the data to some file or database when you terminate the python service?
Very interested in a0. I have been using hermes for few months, started at my windows workstation in wsl2, migrated to run in docker on gb10, finally liberated and run bare metal on gb10 after seeing it behaving well. It grows some personality, keeps own journal, not just automated memories, structures information into karpahys wiki. Why bare metal? Now it basically runs my cluster. Saves ton of time. Build containers, compiles new vllm, deploys, monitors logs. That was the goal as eventually it will run my quant production operations on remote colicated machines. Really good experience.
I feel like agent 0 is built for a different purpose. It’s my like I hate to say it, copilot in office 365 or gemini in Google docs. Productivity agent, not sovereign system operator. But I never tried it, just impression from demos I seen and docs I read
I’ve investigated firstly how Engram was integrated with Codex and replicated this way for Hermes (this step needs some system prompt customization in Hermes). Engram has everything built-in - SQLite + FTS5 full-text search. What does it give to you? Memory across paid and local agents - so my Codex experience of solving hard problems get shared with local models.
I used obsidian with rag embedding but then switched to plain karpathu wiki, I use it and agent uses it. Works well
Thanks for the reply.
I assume that you are talking about
not
My original post was based on the assumption that you were talking about deepseek-ai Engram. I apologize for the confusion
Is it necessary to put a guardrail against agent or accidental modification of karpathy llm wiki?
Git. That is all. :)
Got it. I guest what is left is implement something to backup the Git
Unless you have something insanely private there - use use github private repo. I am switching now to forgejo, I run it on one node and cron-sync to another, but I am just seeing bad things happen at github, otherwise github is the most frictionless option
I use forgejo, and just always tell the agent to push to origin as well as commit.
Started with OpenClaw, then also installed Hermes. Both work fine. Nevertheless I am now rebuilding the skills I use with both with langgraph.
In future I am planning to let Hermes handle the scheduling of the langraph jobs and the storage of the results.
You should try A0. Decided to test it for “real use” today. It’s rather good, especially Docker-native isolation (and browser works lovely). In fact, it’s the only agent I’m not afraid to run, as it is isolated by design, as opposed to Claw etc.
On DS4F it developed slash-autocomplete extension quite confidently. Took about an hour and 3 prompts or so.
Anyone tried Odysseus from Mr. PEWDIEPIE? Ether good reviews so far. BTW Hermes native electron dashboard is pure crap, buggy as hell. TUI use only.