Hermes Agent prompt caching is how Hermes avoids paying full price to re-send your whole conversation on every turn. On 19 November 2026, Hermes v0.23.3 added one controller that runs the cache’s whole life: it keeps the cache warm through long tasks, trims idle chats before the one-hour cache runs out, and sets the cache setting to auto by default. In the video below I show how to stop wasting Claude tokens. Below I’ll explain how Hermes prompt caching works, what changed, and the settings that save you the most.
Key takeaways
- Prompt caching lets Hermes re-use the start of your conversation instead of paying full price again.
- It switches on by itself for Claude models on the Anthropic API or OpenRouter.
prompt_caching.cache_ttl=auto(default),5mor1h.- Switching model mid-chat wipes the cache: the next turn gets zero cache hits.
- Want Hermes set up to save tokens? Join the AI Profit Boardroom.
What is prompt caching in Hermes Agent?
Every time you send Hermes a message, it sends the AI model the whole conversation: the system prompt, its tools and every message so far. Without caching, you pay for all of that again on each turn.
With prompt caching, the model provider remembers the start of the conversation for a while. The next request only pays full price for the new part. Hermes’ caching docs say this cuts input token costs by about 75% on multi-turn conversations.
How Hermes places the cache markers
Anthropic allows up to four cache breakpoints per request. Hermes uses a layout it calls system_and_3: one breakpoint on the system prompt, which stays the same every turn, and a rolling window on the last three messages. The details are in Anthropic’s prompt caching guide.
What changed on 19 November 2026
The v0.23.3 release rolls up about 370 merged pull requests. For caching, the notes list one prompt-cache lifecycle controller with four parts:
- Warming through long tool runs. When Hermes is busy running tools, it keeps the cache alive so the next step doesn’t start cold.
- Compaction before the 1-hour cache expires. If a chat goes idle, Hermes can summarise it before the cache runs out, so you don’t pay to re-read a long history later.
- Nous Portal on the full 1-hour lifecycle.
prompt_caching.cache_ttlnow defaults toauto.
In short: long agent jobs and chats you come back to later should cost less, with no setup from you.
Want Hermes running cheaper and faster in your business? Inside the AI Profit Boardroom you get the Agent OS — Hermes, OpenClaw, OpenCode, Claude and Codex in one place — plus step-by-step Hermes training and four live coaching calls a week.
Want more customers from Google too? Get your Backlink Gameplan: we reverse-engineer why your competitors outrank you and send the exact backlinks to close the gap, within 48 hours, or try Polar Rank free for 3 days.
Join the AI Profit Boardroom →Get Your Backlink Gameplan →Try Polar Rank Free for 3 Days →
The cache_ttl setting: auto, 5m or 1h
You set it in config.yaml:
prompt_caching:
cache_ttl: "auto" # or "5m" or "1h"
| Setting | What it does | Best for |
|---|---|---|
auto (default) |
1 hour for human-paced chats, 5 minutes for machine-paced jobs | Almost everyone |
5m |
Short cache everywhere | Fast, back-to-back automated runs |
1h |
Long cache everywhere | Long chats where you take breaks between messages |
Under auto, Hermes treats sub-agents, cron jobs, webhooks, kanban tasks, API calls and batch runs as machine-paced, so they get the 5-minute cache. Your normal chats get the hour. A one-hour cache costs more to write, so auto only uses it where you’re likely to come back.
How cache warming works
Near the end of the cache’s life (at about five-sixths of the time-to-live), Hermes decides once whether to keep it warm with a small refresh request. The docs set limits so this never runs away:
- At most three idle keep-alives per message you send.
- If the cache misses twice in a row for real, Hermes turns warming off for that session. You’ll see
warming_offin/context. - Starting a new chat with
/newor switching route resets it.
How to save the most tokens with Hermes
- Pick one model and stick with it. The docs warn that a
/modelswitch, a fallback to another model or a credential rotation means the next request gets zero cache hits. - Leave
cache_ttlonautounless you have a clear reason. - Keep your system prompt stable. Changing it mid-chat breaks the cache.
- Use
/newfor new topics instead of dragging a huge chat around. - Check
/contextto see what’s using your context window.
Prompt caching switches on by itself when the model is a Claude model and the provider supports cache markers, such as the native Anthropic API or OpenRouter.
Want Hermes set up with you, step by step? Join the AI Profit Boardroom → for the Agent OS and Hermes training.
My verdict on Hermes Agent prompt caching
You don’t need to do anything to get most of the benefit: caching turns on for Claude models and the 19 November update makes it smarter about long tasks and idle chats. The one habit that matters is staying on one model per conversation. Do that, and Hermes Agent prompt caching quietly cuts your token bill on every long job.
About Julian Goldie
I’m Julian Goldie — founder of Goldie Agency and the AI Profit Boardroom. I test new AI agents almost every day for 428K+ YouTube subscribers and run Hermes, OpenClaw, OpenCode, Claude and Codex side by side in my own Agent OS. Join the AI Profit Boardroom for the daily builds, or Get your Backlink Gameplan if you want your own site to rank, or try Polar Rank free for 3 days.
FAQ: hermes agent prompt caching
Does Hermes Agent support prompt caching?
Yes. It switches on automatically for Claude models on the native Anthropic API or OpenRouter.
What is cache_ttl in Hermes?
The prompt_caching.cache_ttl setting in config.yaml. It can be auto (the default), 5m or 1h.
What does auto mean for the Hermes cache?
A 1-hour cache for human-paced chats and a 5-minute cache for machine-paced jobs like cron, sub-agents and webhooks.
What changed in Hermes prompt caching on 19 November 2026?
v0.23.3 added one cache lifecycle controller: warming through long tool runs, compaction of idle chats before the 1-hour cache expires, and auto as the default setting.
Why did my Hermes cache stop working?
Usually a model switch, a fallback model or a changed system prompt. Each one makes the next request start with zero cache hits.
How much does prompt caching save?
Hermes’ docs say about 75% on input tokens in multi-turn conversations.
Ready to put Hermes to work in your business? Join the AI Profit Boardroom for the Agent OS, setup training and daily updates.
Want your site to rank too? Get your Backlink Gameplan and see the exact links that close the gap on your competitors, or try Polar Rank free for 3 days and let AI agents do your SEO on autopilot.
Starting from zero? Get the free AI course and community.
Join the AI Profit Boardroom →Get Your Backlink Gameplan →Try Polar Rank Free for 3 Days →Get the FREE AI Course + Community →
Related reading
- Claude Code Prompt Cache Metrics
- OpenRouter Response Caching
- Hermes Agent Sandbox
- Best Hermes Agent Plugins
Last updated November 2026. This page is a living guide to hermes agent prompt caching — the facts here move fast and it is updated as they do.
