Hermes Agent prompt caching now offers four settings, and this comparison shows which one to pick for your workload after the 20 November 2026 v0.23.3 update.
The update put one controller in charge of warming, compacting and expiring the cache, and made auto the default, per the release notes.
Below: the settings compared, how providers differ, common mistakes and who each setting suits.
Four days earlier, v0.23.2 had already moved every session onto a single cache clock.
cache_ttl options compared
| Setting | What it does | Best for |
|---|---|---|
| "auto" | 5m for typed chats until the first pause over 5 minutes, then 1h; machine jobs stay 5m | Almost everyone |
| "5m" | Always the short tier | Fast back-and-forth sessions and batch jobs |
| "1h" | Always the long tier | Sessions with long gaps every time |
| "off" | No cache markers | OAuth plans billing cache writes, proxies with their own cache_control |
Source: Hermes configuration docs.
Why auto wins for most users
On Claude a 1-hour write costs 2x base input and a 5-minute write 1.25x (Anthropic docs).
Auto pays the premium only after you have actually paused, so a chat that never pauses never pays for the long tier.
The Hermes docs report roughly 40 percent lower interactive cache-write spend on one install using auto.
Providers compared
| Provider route | Behaviour |
|---|---|
| Claude via Anthropic, OpenRouter, Nous Portal | Marked caching at 5m or 1h |
| Nous Portal | Full 1h lifecycle since v0.23.3 |
| Qwen Cloud | 5-minute cap |
| OpenAI-compatible | Automatic, treated as 30 minutes |
| OpenAI Responses, openai-codex, xAI | Never warmed |
| Local models | No clock expiry |
🔥 Compare model and caching setups with other operators Inside the AI Profit Boardroom you get step-by-step video trainings, the installable Agent OS, four coaching calls a week and 3,400+ members. → Join the AI Profit Boardroom
Warming and compaction at a glance
- Warm at 5/6 of the lifetime with a one-token request.
- Up to 3 idle keep-alives per message you send.
- Unlimited warming during a running tool call, capped by cost.
- Compact idle chats with a window of 30 minutes or more when the context is at least half the threshold.
- Offload old tool results to disk when the cache expires.
Common mistakes
- Pinning "1h" for subagents; they are clamped to 5 minutes anyway.
- Leaving an old "5m" pin in config; the v54 migration only removes the old template line, so check yours.
- Expecting warming on Responses API routes, which are never warmed.
- Ignoring
/context, which shows whether the tier is proven. - Using budget-based extended thinking and expecting keep-alives; those requests are never warmed.
Rank in AI Search or It's FREE →
Who each setting suits
- Auto: solo operators and teams mixing chat and automation.
- 5m: batch pipelines and rapid chats.
- 1h: analysts who return to the same long session every half hour.
- Off: specific billing or proxy setups.
Pair the setting with a model from our best Hermes Agent model roundup.
What changed in v0.23.3, at a glance
| Change | Before | After |
|---|---|---|
| Long tool runs | Cache could expire mid-run | Kept warm until the tool returns |
| Idle chats | Expired cold, full re-send on return | Compacted before the one-hour cache expires |
| Nous Portal | Shorter lifecycle | Full one-hour lifecycle |
cache_ttl default | Fixed tier | auto |
Source: the v0.23.3 release notes, which describe one prompt-cache lifecycle controller.
What breaks the cache, compared
| Action | Cache effect |
|---|---|
| Adding a new message | Prefix kept; rolling window moves forward |
| Editing the system prompt mid-chat | Whole prefix rewritten |
| Inserting or removing earlier messages | Cache lost from that point on |
/model switch, fallback or credential rotation | Zero cache hits on the next request |
| Compression | Compressed region recached; system prompt cache survives |
All rows come from the Hermes developer guide on compression and caching.
Session types compared
| Session source | Pace | Tier under auto |
|---|---|---|
| CLI, TUI, desktop, chat platforms | Human-paced | Can use the one-hour tier |
| Subagents and delegated children | Machine-paced | 5 minutes |
| Cron, webhook, Kanban, API server, batch | Machine-paced | 5 minutes |
| One-shot commands | Machine-paced | 5 minutes |
| Side agents from /background, /btw, /refine | Side agent | 5 minutes |
The one-hour tier costs more to write, so it only pays when turns are spaced out, which is why machine-paced sources stay short.
Cache breakpoint strategy
Anthropic allows four cache_control breakpoints per request.
Hermes places one on the system prompt and three on the last three non-system messages, a strategy the guide names system_and_3.
That keeps the stable part cached for the whole session while the recent turns roll forward.
Compression layers compared
| Layer | Fires at | Purpose |
|---|---|---|
| Agent context compressor | 50 percent of context (default) | Normal context management inside the tool loop |
| Gateway session hygiene | 85 percent of context | Safety net for sessions that grew between turns |
| Cache-edge compaction (new) | When idle keep-alives run out | Shrink the chat before the cache expires |
Join the AI Profit Boardroom →
Who benefits most
- People running long, multi-turn sessions on Claude models.
- Operators whose agents run long tool calls, such as large file or research jobs.
- Anyone who parks chats and returns later in the hour.
Short single questions gain little, because there is little prefix to reuse.
Spec routes compared
| Route spec | How the TTL is set | Refresh on read |
|---|---|---|
| Anthropic-spec (cache markers present) | 5m or 1h as requested, capped by route ceiling | Per provider rules |
| OpenAI-spec (all other cached requests) | Declared 30 minutes | Extends to read time plus 30 minutes |
| OpenAI-spec with 24h retention | Warming deadline 30m, estimated end 24h | Not pruned cold on the clock |
| Local endpoints | No timed expiry | Not applicable |
All of these are described in the developer guide.
Cache confidence levels compared
| Level in /context | Meaning | Hermes behaviour |
|---|---|---|
| Proven | The response named the tier | Warms and compacts at the cache edge |
| Declared | What Hermes requested | Acts on the requested tier |
| Assumed or unknown | No confirmation | Will not warm or compact at the edge |
"Unknown" often follows a model switch, fallback or compaction that moved the cache.
How to verify savings
Anthropic's API reports cache_creation_input_tokens and cache_read_input_tokens on every response, per its prompt caching docs.
On a warm session most input lands in cache reads.
A turn heavy on cache creation means the prefix was rewritten, usually by a model switch or a prompt edit.
Compare a week before and after the update on the same workload for a fair read.
Pinning vs auto: decision table
| Your pattern | Best setting |
|---|---|
| Mixed interactive and automated use | auto |
| Every session has gaps over five minutes | "1h" |
| Rapid machine-paced jobs only | "5m" or auto |
| OAuth plan billing cache writes as extra usage | off |
Proxy that adds its own cache_control | off |
Options and their behaviour are listed in the configuration docs.
Warming rules compared
| State | Warming behaviour |
|---|---|
| Idle chat | At most three keep-alives per message you send |
| Tool running | Keeps warming until the tool returns, bounded by cost |
| Two misses in a row | Warming switches off for the session |
| Budget-based extended thinking | No keep-alives |
Common mistakes
- Switching models mid-session to compare them, which zeroes the cache each time.
- Editing the system prompt or skills during a long chat.
- Pinning one hour for cron or subagent work.
- Reading "unknown" in
/contextas a bug when it follows a fallback. - Leaving caching on behind a proxy that adds its own cache markers.
Before and after scenarios
| Scenario | Before v0.23.3 | After v0.23.3 |
|---|---|---|
| A tool runs for longer than the cache lifetime | Cache could expire; next turn re-sent in full | Kept warm until the tool returns |
| You leave a chat for 40 minutes | Could expire cold | Idle keep-alives, then compaction before expiry |
| You return after the cache expired | Full history re-sent | Old tool results already offloaded, chat smaller |
| Nous Portal session | Shorter lifecycle | Full one-hour lifecycle |
The pattern is the same in each row: fewer full-price re-sends.
Glossary
| Term | Meaning |
|---|---|
| TTL | How long the provider keeps the cached prefix |
| Breakpoint | A cache_control marker telling the provider where to cache |
| Keep-alive | A tiny replay that refreshes the cache before it expires |
| Cache edge | The moment the cache is about to expire |
| Route ceiling | A cap on how long Hermes assumes a provider keeps the cache |
| Offload | Moving old tool results to disk with a pointer left in place |
Quick picks
- Most users: update and leave
cache_ttlonauto. - Long gaps every session: consider pinning
"1h". - Cache writes billed as extra usage: turn caching off.
- Provider keeps less than asked: set a
route_ttl_ceilingsentry. - Local models: nothing to do; no timed expiry.
Options are documented in the configuration docs.
Checklist after updating
- Hermes on v0.23.3 or newer.
cache_ttlonautoor unset./contextshows a warm cache on a Claude session.- No mid-session model switches in your usual workflow.
- Usage logs show most input as cache reads on long sessions.
Session habits compared
| Habit | Cache outcome |
|---|---|
| One model for the whole session | Prefix reused every turn |
| Skills loaded at the start | System prompt stays stable |
New session per topic with /new | Small, relevant prefix |
| Switching models to compare answers | Zero cache hits after each switch |
| Editing skills mid-chat | Prefix rewritten from the change |
| One giant chat for unrelated jobs | Large prefix, more compaction, more rewrites |
Rank in AI Search or It's FREE →
The top three habits are free and matter more than any config change.
Compression and the cache
When Hermes compresses a long chat, the compressed region loses its cache but the system prompt cache survives.
The rolling window of recent messages rebuilds caching within one or two turns, per the developer guide.
So compression costs a little once, then savings resume quickly.
That is why cache-edge compaction is a net win: it trades one small rewrite for avoiding a full cold re-send later.
Who should change nothing
If you use Claude through Anthropic or OpenRouter, run mixed chats and background jobs, and have never touched prompt_caching, you are already on the best setting after the update.
The only action is to update Hermes and keep the session habits above.
Where to look when costs rise
- Run
/contextand check whether the cache shows warm and proven. - Check whether a model switch, fallback or credential rotation happened in the session.
- Check whether skills or the system prompt changed mid-chat.
- Check whether the session is machine-paced and therefore on the five-minute tier.
One of those four explains most cost spikes.
Verdict
Stay on auto, read /context once a week, and learn the wider agent stack in the AI Profit Boardroom.
Join the AI Profit Boardroom →
FAQ: Hermes Agent Prompt Caching Compared
Is auto better than 1h?
For most users yes, because it pays the 1-hour premium only after you pause.
Do subagents use the 1-hour cache?
No, delegated subagents are always clamped to 5 minutes.
Does OpenAI get warmed?
OpenAI-compatible routes are warmed on a 30-minute assumption; Responses API sessions are never warmed.
Where do I check the tier in use?
Run /context in the session.
📺 Video notes + links to the tools 👉
🎥 Learn how I make these videos 👉
🆓 Get a FREE AI Course + Community + 1,000 AI Agents 👉
About Julian
I'm Julian Goldie, an SEO entrepreneur, author and founder of the AI Profit Boardroom, which has 3,400+ members.
I help business owners scale with AI agents, automation and SEO.
- I've built a 7-figure SEO and link building agency, Goldie Agency, with a team of 70+ SEO experts.
- I've grown a YouTube channel to 400,000+ subscribers.
- I wrote "Link Building Mastery".
- My Udemy courses have taught over 29,000 students.
→ Get my best AI training inside the AI Profit Boardroom
Check your config.yaml for an old pin, set auto if you find one, and compare costs after a week.
For guided setups, join the AI Profit Boardroom, and keep this comparison for every Hermes Agent prompt caching decision.











