Is Hermes Compaction Worth Setting Up For Your Business?

Hermes compaction is what lets one business agent hold a week-long customer thread, a multi-day job or a busy client channel without running out of memory mid-conversation.

When a chat grows past a set share of the model's window, Hermes Agent writes a structured summary of the older middle and keeps the newest messages exactly as they were.

On 2 November 2026, release v0.22.6 added an opt-in setting, compression.prepare_ahead, that writes that summary early so the agent has it ready.

This page is for owners who want to know if any of that is worth an afternoon of setup.

If you want a second opinion on where an agent fits your business, book a free SEO strategy session with my team.

What Hermes Compaction Actually Does For A Business Agent

Every AI model can only read so much at once.

A support bot that has answered forty questions, read ten files and run a dozen tools hits that limit sooner than most owners expect.

Hermes handles it for you: the official configuration page says it compresses long conversations automatically, and the default trigger is half of the model's window.

Your agent keeps the goal, the decisions, the files it touched and the next steps in the summary.

The bulky stuff, like old terminal dumps and long tool results, gets cleared first because it costs the most and matters least later.

You do not have to switch any of this on; it is enabled out of the box.

Why Long Chats Get Expensive And Slow Without It

An agent re-sends its history to the model on every turn.

If nobody trims it, each reply carries more tokens than the last, which means slower answers and bigger model bills.

For a cost ceiling, the docs offer threshold_tokens, an absolute cap on when compaction fires.

Their own example compacts a 1M-window model at 256K tokens rather than waiting until 500K.

That is the one setting I would point an owner at first, because it turns an open-ended cost into a predictable one.

There is a second, cheaper lever too: proactive_prune_tokens trims old tool output with no model call at all, and the docs suggest 48000 as a starting value.

Use Case: A Local Shop's Telegram Order Bot

Picture a bakery that takes custom cake orders through a Telegram bot running on Hermes.

A customer messages on Monday, goes quiet, then returns on Thursday to change the icing.

Without help, the bot would re-read the whole stale thread on each new reply.

The docs fix that with idle_compact_after_seconds: set it to 1800 and a thread that wakes after 30 quiet minutes gets compacted before the first answer.

The customer sees none of this, because automatic compaction stays silent on chat apps unless you set progress_notices: true.

A failed compaction still shows a notice, so you are never left guessing when something breaks.

Use Case: A Service Firm Running Multi-Day Jobs

Now take a bookkeeping firm whose agent works through a client's receipts over three days in one session.

The worry here is detail loss: did the agent forget the client said "don't touch the 2025 folder"?

The default tail mode, called lean, is built for exactly that, and the developer guide says it quotes every real user message in the summary.

It also keeps an index of paths, IDs and error strings pulled out word for word, plus a pointer the agent can use to search back through anything summarised.

Old turns are not deleted either.

With in_place left at its default of true, the session keeps one ID and older messages are archived under it, still searchable.

🔥 Want an agent that handles long client threads without you babysitting it? 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 building agents for real businesses. → Join the AI Profit Boardroom

Use Case: An Agency With A Different Model Per Client

Agencies often swap models: a big-window model for research-heavy clients, a smaller one for quick social replies.

Hermes lets the trigger follow the model with model_thresholds, so a 1M-context model can compact later and a 128K model earlier.

Keys match part of the model name, and the longest match wins.

For agencies that hand work to subagents, the delegation docs say children compact at the same point as their parent unless you set delegation.compression_threshold_tokens, which is off by default.

That keeps one rule across the team instead of a dozen hand-tuned bots.

It also makes billing conversations easier, because you can tell a client roughly when a long session will be condensed instead of guessing.

The pressure bar in the command line helps here: it shows how close a session is to the trigger, with a first warning at 60% and a stronger one at 85%.

prepare_ahead, Explained Without The Jargon

Normally the summary gets written at the moment compaction fires, and the conversation pauses while it happens.

With prepare_ahead: true, Hermes starts writing it a little earlier, in the background, while tools are running or between replies.

When the trigger finally hits, the ready-made summary drops in if it still covers the same messages.

Anything newer stays untouched in the recent tail.

The change came in commit 32a636bc8d and is part of the v0.22.6 release notes as "opt-in prepared compaction".

It is off by default, and you flip it with one line under compression in the config file.

Where The Time Saving Shows Up, And Where It Does Not

The saving is the pause you no longer sit through at the moment of compaction.

That helps agents that stay alive across many turns: the interactive CLI, the Desktop app and messaging bots on Telegram, Discord or Slack.

It does nothing for scheduled cron jobs, delegated helpers or one-shot commands, because those agents finish with the turn and skip the end-of-turn pass.

It also does nothing when someone types /compress by hand, or when you run a plugin context engine instead of the built-in one.

If your bots mostly run on cron, leave the setting off and put your hour into something else.

The Real Risks: Lost Detail And Wasted Calls

Compaction is lossy by design.

The built-in engine summarises; it does not keep every word, so fine detail from early in a long job turns second-hand.

The biggest avoidable risk is picking a summary model with a smaller window than your main model.

The developer guide warns that the call then fails and the middle of the chat is dropped with no summary at all.

prepare_ahead adds its own small cost: the docs say it spends an extra summary call that may end up unused if the conversation changes before the trigger.

And if your memory provider feeds context into summaries, it switches itself off after the first compaction.

Pick the summary model on purpose, because every summary is a real call to whichever provider you point it at.

What Your Agent Still Knows After A Compaction

Owners usually ask one question: will the bot forget what the customer wanted?

The summary is written to a fixed outline with sections for the goal, the customer's preferences and constraints, finished and unfinished work, key decisions, files and next steps.

The opening exchange is pinned too: by default the first three non-system messages survive every pass, so the original request stays in view.

If the agent had loaded a skill, such as your refund policy playbook, and that skill gets summarised away, Hermes marks it and tells the agent to reload it before relying on it.

That is a sensible safety net for any business that runs agents on written procedures.

What gets thinner is colour: long tool outputs and step-by-step narration become a line or two.

For most customer and client work, that trade is fine.

Micro-Compaction: Smoother Chats, Weaker Cache

There is a third style, described on the micro-compaction page.

Instead of one big summary, Hermes folds one old exchange into a running summary after every finished turn.

Owners like the idea because there is never one long stall.

The cost is that the model provider's prompt cache breaks every turn, and the docs say plainly that for some setups this costs more than the stall it removes.

It is off by default, and I would only try it after measuring your own sessions.

📞 Not sure if your business needs a long-running agent at all? Book a free SEO strategy session and my team will look at your workflow and where an agent actually saves hours. → Book your free call

What To Ask Your Team Or Your Agent To Set

Keep compression.enabled on and the default trigger of 0.50.
Pick a summary model under auxiliary.compression whose window matches or beats the main model.
Set threshold_tokens if you want a fixed ceiling on each call.
Add idle_compact_after_seconds: 1800 for customer chat bots.
Turn on prepare_ahead for agents that live in Desktop, the CLI or a chat app.
Leave micro-compaction off until you have numbers.

Your team can change these keys on a live gateway, and the docs say the change applies on the next message with no restart.

If your bots also stream replies into chat apps, our Hermes streaming guide for business bots covers the stall caps that sit next to these settings.

Getting v0.22.6 Onto Your Machines

The release notes are clear on who gets the update when.

Anyone on the command line runs hermes update.
The Microsoft Store app updates itself through the Store.
Desktop copies installed from the .exe or .dmg wait for the next bundled release.
Docker and Hermes Cloud users pull nousresearch/hermes-agent:stable.

If staff run the Desktop installer build, expect prepare_ahead to arrive a little later on their machines.

Older config files are safe: the docs say legacy summary-model keys move to the new auxiliary block automatically the first time Hermes loads them.

My Call For Small Businesses

Leave the defaults on, add a token cap, and pick the summary model with care; that covers most of the value.

Turn on prepare_ahead only for agents that hold long, live conversations, and skip it for scheduled jobs.

Treat micro-compaction as an experiment, not a default.

If you want the trainings and the ready-built Agent OS to run all of this, join the AI Profit Boardroom.

Also On Our Network

FAQ: Is Hermes Compaction Worth Setting Up For Your Business? Use Cases, Risks And The New prepare_ahead

Do I have to turn compaction on?

No. Compression is enabled by default in Hermes Agent and fires at half of the model's context window unless you change the threshold.

Will customers see a message when the bot compacts?

Not by default. Routine compaction is silent on chat platforms unless progress_notices is set to true; failure notices still show.

Does prepare_ahead help scheduled jobs?

No. Turn-end passes are skipped for cron runs, delegated children and one-shot sessions because those agents end with the turn.

Can compaction cut my model spend?

It keeps re-sent history smaller, and threshold_tokens gives each call a fixed ceiling. prepare_ahead adds one extra summary call that may go unused.

Is old conversation history deleted?

No. With in_place left on, older turns are soft-archived under the same session ID and stay searchable through session_search.

📺 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

Start with one customer-facing bot, watch a week of long threads, and if you want my team to sanity-check the plan, book a free SEO strategy session before you roll out Hermes compaction across the business.