Hermes Agent's Claude prompt cache isn't "all conversations are automatically half price". It primarily caches stable prefix content such as system prompts, skills, memories, and repeating contexts; If you frequently change models, providers, profiles, or stuff new large files every round, the caching benefits will become smaller.
The official description for v0.14 mentions that Hermes Agent adds a cross-session 1-hour prompt prefix cache to Claude on the Anthropic, OpenRouter, and Nous Portal paths. The focus is on "prefix": the more consistent the stable content in the previous paragraph, the easier it is to hit; Later user new inputs and tool results will still be billed normally.
Why do you feel unconscious
- There is no hot cache when you first use it, and the first round is usually not immediately cheap.
- Changed the profile, skills, and memory contents, and the prefix changed.
- Switching from Anthropic to OpenRouter or Nous Portal, cache links are not necessarily shared.
- Upload new files and paste new logs every time, and the main cost comes from new content.
- The model or provider itself is not exposed to cache discounts as you would expect.
So don't think of prompt cache as a one-size-fits-all money-saving switch, it's more like an "accelerator when using the same set of work contexts for a long time".
How to improve the hit rate
The most effective approach is to reduce meaningless contextual jitter: use the same Claude provider steadily; Do not modify system prompts or large chunks of memory every round; Put project rules into fixed files or skills; Large files are only re-read when necessary; Short tasks use ordinary conversations, and long tasks make the cache more valuable.
If you mainly do one-time Q&A, the caching income is limited; If you work around the same codebase, the same set of skills, and the same profile every day, 1 hour of cross-session caching is easier to reflect. If you want to check the feature boundaries, you can check the official Hermes Agent repository release:https://github.com/NousResearch/hermes-agent.