Hermes Agent prompt caching just got smarter. In the 19 November 2026 canary and the v0.23.3 release on 20 November, Hermes moved to one prompt-cache controller that keeps the cache warm through long tool runs, trims idle chats before the 1-hour cache runs out, and sets the prompt_caching.cache_ttl setting to auto by default. In plain words: long agent jobs re-read less of your conversation at full price, so they cost less and start replies faster. This guide explains how it works and how to get the most from it.
📺 Watch: Hermes Agent Just Got 10X Faster!
🔥 Get the Agent OS as a free bonus: AI Profit Boardroom members get the full Agent OS zip, prompt libraries, daily tutorials and weekly live coaching calls. → Get inside
What Prompt Caching Is (in Plain English)
Every time Hermes sends a message to an AI model, it sends the whole conversation again: the system prompt, the tool list, every past message and every tool result. On a long task that can be a lot of text, and you pay for all of it each turn.
Prompt caching lets the AI provider remember the start of that conversation for a while. When the next request begins with the same text, the provider reads it from the cache instead of processing it fresh. Cached reads are much cheaper and quicker.
The Hermes docs say prompt caching cuts input token costs by about 75% on multi-turn conversations with Anthropic models. That's the docs' figure for Claude models. Your own savings depend on the model, the provider and how you work.
Want a Hermes setup that is already tuned for cost? Inside the AI Profit Boardroom we share our Agent OS configs and walk through them on weekly live calls.
What Changed in the Hermes Agent Prompt Caching Update
The v0.23.3 release notes list these changes as part of the new prompt-cache lifecycle controller:
- Warming through long tool runs. When Hermes is busy running tools, it can now keep the cache alive so the next model call still gets cheap cached reads.
- Compaction of idle chats before the 1-hour cache expires. If a chat goes quiet, Hermes can summarize it while the cache is still warm, instead of paying full price to re-read everything later.
- Nous Portal on the full 1-hour lifecycle.
prompt_caching.cache_ttldefaults toauto.
The 19 November canary notes add some detail: the cache-expiry compaction summarizes using the conversation's warm prefix, and the warm summary "runs only when cheaper". Large chats also need to grow again before they are re-compacted.
📺 Watch: Hermes Agent OS Can Automate EVERYTHING?! 🤯
How the Cache Setting Works
You control it in config.yaml:
prompt_caching:
cache_ttl: "auto"
There are three options, according to the docs:
- auto (default): 1 hour for human-paced sessions, where you take breaks between messages, and 5 minutes for machine-paced ones.
- 5m: use the 5-minute cache everywhere.
- 1h: use the 1-hour cache everywhere.
The docs list the machine-paced sources that get 5 minutes under auto: subagents, cron jobs, one-shot runs, webhooks, kanban, API and API server calls, tools, batch jobs and the curator. Side agents such as /background, /btw and /refine also get 5 minutes.
Why does this matter? A 1-hour cache costs more to write than a 5-minute one, but it survives your coffee break. A cron job that fires and finishes fast doesn't need an hour. Auto picks for you, so most people should leave it alone.
How Hermes Keeps the Cache Warm
The developer docs explain the warming logic. Here's the short version:
- Near the end of the cache's life (at about five-sixths of the time-to-live), Hermes decides once whether to send a small keep-alive request.
- It sends at most three idle keep-alives per user message.
- While tools are running, keep-alives are booked on that tool run instead, limited by a cost check.
- If keep-alives miss the cache twice in a row for real, Hermes switches warming off for that route for the rest of the session, so it doesn't keep paying for nothing.
- If warming isn't worth it, Hermes can compact the chat, or mark the old history as cold and offload it.
The goal is simple: only spend money on warming when it saves more than it costs.
We test which models give the best results for the money on Goldie Bench, and our free Agent OS guide shows how we run long Hermes jobs day to day.
📺 Watch: Stop Finding Leads Manually! Use Hermes Agent
Which Models Use It
The docs say prompt caching turns on automatically when the model is an Anthropic Claude model and the provider supports cache markers, such as the native Anthropic API or OpenRouter. Hermes uses a "system_and_3" strategy: one cache marker on the system prompt and a rolling window on the last three messages, inside Anthropic's limit of four markers per request.
The docs also describe how Hermes tracks caches for other providers that cache automatically (it treats those as a 30-minute cache by default), and how Nous Portal can pick which models get cache markers from its model catalog.
Speed and Cost: What You Actually Get
Here's what this means for a normal business user running Hermes:
- Long research or coding jobs cost less. The big system prompt and tool list are read from cache instead of re-sent at full price each step.
- Coming back after a break is cheaper. With a 1-hour cache and idle compaction, you're less likely to pay to re-read a huge chat from scratch.
- Replies can start faster. Cached text doesn't need to be processed again, which helps with long conversations.
- Background jobs don't overpay. Cron and subagent runs use the shorter cache automatically.
Want help setting up cron jobs and subagents that use this well? The AI Profit Boardroom has step-by-step tutorials and prompt libraries.
5 Habits That Keep Your Cache Hits High
- Don't switch models mid-chat. The docs are clear: caches are tied to the model and account. A
/modelswitch, a fallback to another model or a credential rotation means the next request gets zero cache hits and re-reads the whole chat at full price. - Keep your system prompt stable. Changing it mid-conversation breaks the cache for everything after it.
- Use
/newfor a new topic. A fresh chat is cheaper than dragging a huge old one along. - Leave cache_ttl on auto unless you have a clear reason to pin it.
- Update Hermes. These cache changes landed on 19 and 20 November 2026, so older versions won't have them.
How to Check It's Working
The docs mention that when warming is switched off for a route, the /context view shows it as warming_off. Use /context during a long session to see what's going on. If you see warming off a lot, check whether something keeps changing your model or provider mid-chat.
For more on keeping long sessions lean, see our guides on Hermes memory management and the Hermes Agent fallback model. A fallback switch can cost you your cache, so set it up with care.
FAQ
Is Hermes Agent prompt caching on by default?
Yes, for Claude models on providers that support cache markers, such as the Anthropic API and OpenRouter. The cache_ttl setting defaults to auto.
What does cache_ttl auto do?
It picks a 1-hour cache for human-paced chats and a 5-minute cache for machine-paced work such as cron jobs, subagents and webhooks.
How much does prompt caching save?
The Hermes docs say it cuts input token costs by about 75% on multi-turn conversations with Anthropic models. Real savings vary by model and workflow.
Why did my cache stop working?
The most common cause is a model change mid-chat, from a /model switch, a fallback or a credential rotation. Each one starts the cache from zero.
Which version added the new cache controller?
It appeared in the 19 November 2026 canary and is listed in the v0.23.3 release notes from 20 November 2026.











