In this guide to Hermes Agent prompt caching, here’s what we’re going to cover: why it matters, when it switches itself on, and how to stop your cache from quietly resetting. Think of it as a short-term memory on the provider’s side. Hermes tells the model which opening chunk of your chat will stay the same, the provider stores that chunk, and later turns skip re-reading it. I recorded the video below on cutting wasted Claude tokens, and everything underneath is checked against the official Hermes docs and the latest release notes.
The short version
The docs credit caching with cutting input token costs by about 75% on multi-turn chats.
You get it automatically on Claude models through Anthropic or OpenRouter, with Nous Portal choosing via its catalogue.
Hermes splits four breakpoints into one for the system prompt and three for recent messages.
The TTL default is auto: an hour for people, five minutes for automated sources.
Agents are chatty. A single Hermes session can run for dozens of turns, and each turn re-sends the system prompt, the tools and the full history to the model. As the conversation grows, so does the pile of text the model has to chew through before it reaches your newest message.
Caching takes the repeated part out of that pile. According to the context compression and caching docs, it reduces input token costs by around 75% on multi-turn conversations. That is the single biggest lever on how much a long Hermes session costs to run, and it needs no effort from you once the conditions are right.
Which models and providers get the cache
Hermes switches caching on when the model is an Anthropic Claude model and the provider supports cache_control markers. Today that covers two routes:
Route
Cache markers?
Notes
Native Anthropic API
Yes
Tool messages get markers here too
OpenRouter
Yes
Marker caching for Claude models
Nous Portal
Model by model
The catalogue decides which models take markers
If you’re on another route, the provider’s own caching applies instead and Hermes still keeps one cache clock for the session. The breakpoint behaviour explained below belongs to the Claude route.
Want your Hermes sessions set up properly from day one? Members of the AI Profit Boardroom get the Agent OS (Hermes, OpenClaw, OpenCode, Claude and Codex together), guided Hermes lessons and four live coaching calls every week.
Anthropic lets a request carry no more than four cache breakpoints, so Hermes has to choose where they go. Its answer is called system_and_3.
Breakpoint one sits on the system prompt, the most stable part of any request.
Breakpoints two to four form a rolling window over the three most recent non-system messages.
The markers themselves are tiny: {"type":"ephemeral"} for the standard cache, with "ttl":"1h" added for the long one. Hermes injects them through apply_anthropic_cache_control(), which works on a deep copy, so your stored messages stay untouched. You can read how breakpoints behave on the provider end in Anthropic’s documentation.
How to tune the cache TTL
The TTL is how long a cached prefix lives before the provider drops it. Hermes reads it from config.yaml under prompt_caching:
With auto, Hermes picks per source. Sessions a human is typing into get the one-hour cache. Machine-paced sources get five minutes, and the docs name them: subagent, cron, oneshot, webhook, kanban, api, api_server, tool, batch and curator. Side agents from /background, /btw and /refine sit on five minutes as well.
My advice is simple. Stay on auto unless you know your pattern. If you leave long gaps between messages all day, 1h suits you. If everything you run is automated and rapid-fire, 5m is the tidier fit.
Five things that wipe your cache
A cache hit needs the start of the new request to match the stored prefix exactly. Change anything near the front and the match is gone from that point on. These five cause most misses:
Cause
What happens
How to avoid it
Model switch via /model
The model is part of the cache key, so the next turn starts from nothing
Pick your model before a long chat and keep it
Fallback to a backup model
Same effect as a manual switch
Keep your primary model healthy and within limits
Credential rotation to another account
The other account holds no cached copy
Expect one cold turn after a rotation
Editing old messages mid-chat
Every message after the edit loses its cache
Add new messages instead of rewriting earlier ones
Idle past the TTL
The stored prefix expires
Use the 1h tier for slow, human-paced work
The first three are the expensive ones: the docs say the next request gets zero cache hits and re-reads the entire conversation at the undiscounted rate.
Fancy building your Hermes setup alongside other members? Hop into the AI Profit Boardroom → for the Agent OS and live help.
Hermes v0.23.3 landed on 20 November 2026 with a single controller that manages the whole cache lifecycle. According to the release notes, it now:
keeps the cache warm during long tool runs;
compacts idle chats before the one-hour cache runs out;
puts Nous Portal on the full one-hour lifecycle;
sets prompt_caching.cache_ttl to auto by default.
The 19 November canary filled in the finer points. Compaction at cache expiry summarises from the warm prefix, that summary creates no fresh cache entry and only runs when it works out cheaper, and auto keeps each source’s tier while clamping per send. A broken prompt_caching section in your config can no longer stop Hermes from starting. One release earlier, v0.23.2 gave each session a single prompt-cache clock.
Checking cache health with /context
Warming is the keep-alive that stops a cache expiring while you think. Hermes makes one decision at around five-sixths of the TTL and caps it at three idle keep-alives per user message, so it can’t run on forever.
Open /context to check the state. A warming_off flag means two genuine misses happened and warming has been switched off for this session; /new gives you a clean slate.
Compression and caching also work together rather than against each other. Compressing a long chat throws away the cache for the squashed section, yet the system prompt cache stays intact and the three-message window recovers within a turn or two. The note compression adds to the system prompt only appears on the first compaction, which keeps that front section steady.
The bottom line
Caching is already doing the heavy lifting if you run Claude through Anthropic or OpenRouter. Your job is to protect the prefix: one model per conversation, a system prompt you don’t fiddle with, and the auto TTL unless your pattern says otherwise. Get those right and Hermes Agent prompt caching keeps long sessions far lighter than they would otherwise be.
About Julian Goldie
Julian Goldie here: I founded Goldie Agency and the AI Profit Boardroom. My YouTube channel has 428K+ subscribers, and most days I’m testing a new AI agent, with Hermes sitting beside OpenClaw, OpenCode, Claude and Codex in my Agent OS. You can follow the builds inside the AI Profit Boardroom, and if your website needs more traffic, get your Backlink Gameplan, or try Polar Rank free for 3 days.
FAQ: hermes agent prompt caching
Why is my Hermes cache missing every turn?
Look for a changing prefix. Switching model, a fallback, a credential rotation, editing earlier messages or long idle gaps past the TTL all reset the match. Run /context; if it shows warming_off, two genuine misses happened, and /new starts a fresh session.
Should I set cache_ttl to 1h or 5m?
Most people should keep auto, which already gives typed chats an hour and automated sources five minutes. Choose 1h if you work slowly with long pauses, and 5m if every run is automated and back to back.
How does context compression affect the cache?
Compression drops the cache for the section it squashes, but the system prompt cache survives and the rolling window over the last three messages rebuilds within one or two turns.
Is Hermes prompt caching automatic?
Yes, for Anthropic Claude models on a provider that supports cache_control markers, such as the native Anthropic API or OpenRouter. Nous Portal uses its model catalogue to decide.
What did v0.23.3 change about caching?
It added one prompt-cache lifecycle controller: warming through long tool runs, compaction of idle chats before the one-hour cache expires, Nous Portal on the full one-hour lifecycle and auto as the default cache_ttl.
How many cache breakpoints does Hermes use?
Four, which is Anthropic’s limit per request: one on the system prompt and three on the most recent non-system messages.
Want Hermes handling real tasks in your business this week? The AI Profit Boardroom gives you the Agent OS, setup lessons and fresh updates whenever Hermes ships.