User reports Hermes Agent got less autonomous after cost tweaks
When the user first installed , it worked impressively well with only minimal prompting, acting to find information and complete tasks on its own. For example, for a service where an API or bots couldn't be used, the user roughly described the goal and Hermes independently found a way to recreate a communication dashboard inside Discord using email, handling the whole solution itself.
Later, while reviewing usage, the user noticed a large cost spike caused by inefficient model usage and heavy consumption. After attempting to make the setup more cost-efficient, things started going downhill.
Now the agent remembers things only within a single session but fails to retain them in its longer-term memory, forgets which skills and tools it has access to, and often announces it will do something but then stays inactive instead of following through. The user isn't sure whether they misconfigured something or whether this is just normal behavior over time.
Key points
- Initially, minimal prompting was enough for Hermes to build a working solution, like an email-based Discord communication dashboard
- The user discovered a big cost spike from inefficient model usage and high volume while reviewing usage
- seemed to decline after the user tried to make the setup more cost-efficient
- The agent now retains information only within a session, not in longer-term memory, and forgets which skills/tools it has
- It repeatedly announces intended actions but then remains inactive instead of completing them