Hermes compaction earns an 8 out of 10 from me: the automatic summariser, the lean uncompressed tail and the failure handling are excellent, and the new prepare_ahead option is useful but narrow.
That rating covers the whole system as the official docs describe it, from the 50% trigger to manual /compress, micro-compaction and plugin context engines.
The ratings below are editorial judgement drawn from the Hermes docs and the v0.22.6 release notes, published 2 November 2026, not from benchmarks.
If you want my team to weigh in on your own agent stack, book a free SEO strategy session.
The Hermes Compaction Scorecard
| Criterion | Score /10 | Why |
|---|---|---|
| Automatic trigger | 9 | Ratio trigger at 0.50, an absolute threshold_tokens cap, and per-model overrides. |
| Summary structure | 8 | A fixed template (Goal, Progress, Key Decisions, Next Steps and more) that is updated, not rewritten, on later passes. |
| Uncompressed tail | 9 | Lean mode keeps a clamped 10K–25K tail and quotes every real user message in the summary. |
| Manual /compress | 8 | Runs on demand, clears the failure cooldown and always reports back on chat apps. |
| prepare_ahead | 7 | Removes the wait at the trigger, but skips manual, plugin, native and short-lived sessions. |
| Micro-compaction | 6 | No single long stall, but the prompt cache breaks on every pass. |
| Plugin context engines | 7 | A clean swap point, never auto-activated, but the built-in extras stop working. |
| Failure handling | 9 | Escalating cooldowns, a fallback chain and a deterministic summary instead of a dead session. |
| Documentation | 8 | Unusually deep, with a couple of small inconsistencies noted below. |
The average lands at 7.9, which I round to the headline 8.
Strength: The Lean Tail Is The Best Part
The default tail_mode: lean is where the context compression developer guide shines.
It keeps 2.5% of the window verbatim, floored at 10K tokens and capped at 25K, and never more than a fifth of the window.
Continuity rides in the summary instead: an identifier-preserving session log, a regex-extracted anchor index and a recovery pointer for session_search.
The guide reports roughly 49K tokens retained against roughly 162K for legacy mode on 500K-token real sessions, with higher recall when recovery is used.
That is a rare case of a smaller context and better memory at once.
Strength: It Fails Soft, Not Hard
Summary backends go down, and Hermes plans for it.
A failed or stalled attempt arms a per-session cooldown that escalates from 60 to 300 to 900 seconds, so a broken summariser is not hammered every turn.
A second stall drops to a deterministic fallback summary rather than leaving the session stuck, and an overloaded provider aborts with the transcript kept intact.
The configuration reference also bounds the wait: context_timeout_seconds defaults to 120 seconds of silence, and context_total_ceiling_seconds to 600.
For a tool people leave running in Telegram for weeks, this is the feature that matters most.
Strength: One Session, One ID
With in_place: true as the default, compaction no longer spawns a chain of renamed sessions.
Older turns are soft-archived under the same ID and stay searchable.
The docs credit this with ending a cluster of bugs around lost goal state and orphaned sessions, and I count that as a real quality-of-life win.
Strength: The Summary Template Keeps Work Moving
A lot of agent tools summarise into free text; Hermes uses a fixed outline with Goal, Progress split into Done, In Progress and Blocked, Key Decisions, Relevant Files, Next Steps and Critical Context.
On later passes the previous summary is updated rather than replaced, so a task moves from In Progress to Done instead of vanishing.
Skills that get summarised away are flagged, and the summary tells the model to reload them, which closes a gap I see in simpler tools.
The budget scales too: a fifth of the content being compressed, with a floor of 2,000 tokens and a ceiling of 12,000 or 5% of the window, whichever is lower.
🔥 Want the compaction settings I rate highest, already wired into an agent? Inside the AI Profit Boardroom you get step-by-step video trainings, the installable Agent OS with Hermes inside, four coaching calls a week and 3,400+ members comparing setups. → Join the AI Profit Boardroom
prepare_ahead Rated: 7 Out Of 10
The headline addition in the v0.22.6 release is opt-in prepared compaction, off by default.
When the provider-reported prompt is within 12% below the trigger, Hermes runs one background summary pass on a copy of the window, leaving the live transcript and prompt cache alone.
At the next automatic compaction that summary is spliced in if the same messages are still there, and newer messages land in the tail.
If the pass is still running, compaction waits for it rather than paying twice.
The design is careful: it checks durable message IDs, tolerates plugin text edits, treats a rewind as a miss and cancels a pass that a session switch overtakes, per commit 32a636bc8d and the developer guide.
A failed pass also pauses further passes without arming the live failure cooldown, so a bad background attempt cannot block normal compaction.
Opt-in shared metrics count every pass and splice outcome under hermes.prepared_compaction.count, which means you can check whether it is actually paying off.
It loses points on reach, not on quality.
Limit: prepare_ahead Skips A Lot Of Sessions
The exclusions list is long enough to shape who benefits.
- Manual
/compress, focus compaction and provider-overflow recovery always run inline. - Plugin context engines, Codex app-server sessions, native Responses compaction and micro-compaction are never prepared.
- Cron runs, delegated children and one-shot
hermes chat -qor-zsessions skip the turn-end pass. - A
-q --resumechain gets no head start, since it builds a new agent each run. - With a memory provider that adds summary context, it turns itself off after the first compaction.
It also costs an auxiliary summary call that may never be used.
For a long interactive CLI, Desktop or chat-app session it is a clear yes; for batch and scheduled work it is a no-op.
Limit: The Summary Model Can Sink It
The single biggest weakness is a configuration trap rather than a code bug.
Per the developer guide, a summariser with less window than the chat model cannot take the single request that carries the whole middle.
Get that wrong and the middle turns are dropped without a summary, which the guide names as the most common cause of degraded compaction quality.
A louder warning at config time would push the summary score higher.
Micro-Compaction Rated: 6 Out Of 10
The micro-compaction docs describe a per-turn mode that folds the oldest unabsorbed exchange into one rolling summary.
Two design choices are excellent: user messages are never summarised, and the rolling summary defrags once it crosses 2,000 tokens.
The problem is cost: each pass rewrites already-sent history, so the provider cache breaks every turn, and the docs themselves call that the strongest argument against turning it on.
It also adds a model call at the end of every turn, with one quoted example of about 31 seconds per pass on a local 7B model.
It ships off, and that is the right default.
Plugin Context Engines Rated: 7 Out Of 10
The context engine setting lets a plugin such as LCM take over with lossless context management by setting context.engine to its name.
Plugin engines are never auto-activated, which I like.
The trade is that built-in extras such as the proactive tool-result prune and prepare_ahead stop applying once a plugin is in charge.
📞 Want a straight answer on whether your agents need any of this? Book a free SEO strategy session and my team will review what you run and where an agent earns its keep. → Get your free strategy session
The Small Extras Rated: Idle Compaction, Notices And Prune
Three opt-in keys round out the system, and each is worth a short score.
idle_compact_after_seconds gets an 8: it compacts a thread that wakes after a long quiet spell, and it refuses to fire when the context is already small.
progress_notices gets an 8 for a sensible default, since chat users see nothing routine unless you ask, while failures and manual results always show.
proactive_prune_tokens gets a 9: a no-LLM trim of bulky tool output with a minimum reclaim of 4,096 tokens so cache breaks stay rare, though it only runs under the built-in engine.
All three can be edited on a live gateway and apply on the next message.
Who Should Use What
Keep Hermes compaction defaults as they are
- Anyone running short to medium chats: the 0.50 trigger and lean tail already handle it.
- Teams on big-window models who want a fixed ceiling: add
threshold_tokensand stop there. - Chat-app bots that sleep for hours: add
idle_compact_after_seconds, with 1800 as the documented example.
Turn on prepare_ahead, or leave it off
Turn it on if you sit in long interactive sessions in the CLI, the TUI, Desktop or a messaging app and you notice the pause when compaction fires.
Skip it if your work is mostly cron jobs, delegated subagents or one-shot commands, if you run a plugin context engine, or if your memory provider writes into summaries.
Skip micro-compaction too unless you have measured your cache discount and decided a flat context is worth more than an intact prefix.
Doc Inconsistencies Worth Knowing
Two small gaps showed up while I compared pages.
The configuration reference treats protect_first_n as a setting you can drop to 0, while the developer guide's parameter table labels it hardcoded.
The developer guide's parameter table also mentions a 0.75 floor for windows below 512K, while its worked example for a 200K model still computes the trigger at 0.50.
Neither changes the verdict, but test on your own model before you trust a number.
For the streaming timers that sit beside these settings, see our Hermes agent streaming verdict.
Final Verdict
Hermes compaction is one of the strongest parts of Hermes Agent: careful trigger logic, a tail that keeps what you said, and failure handling that keeps sessions alive.
prepare_ahead is a well-built 7 that helps long live sessions and nobody else, and micro-compaction is a niche 6.
Update with hermes update, keep the defaults, pick a summary model with a big enough window, and only then try the opt-in extras.
To get the trainings and the Agent OS that runs it, join the AI Profit Boardroom.
Also On Our Network
- 🌐 The eight compaction settings I set before a long Hermes session
- 🌐 Whether compaction is worth it for shops, service firms and agencies
- 🌐 Five prompts that let your agent tune its own compaction
- 🌐 Which benchmarked models to pair as main and summary model
- 🌐 The complete Hermes compaction guide on the AI Profit Boardroom blog
FAQ: Hermes Compaction Verdict
What score does Hermes compaction get overall?
8 out of 10, averaged from a nine-criterion scorecard built from the official docs and release notes.
Is prepare_ahead worth turning on?
Yes for long interactive CLI, Desktop or chat-app sessions. No for cron jobs, delegated children, one-shot commands or plugin context engines, where it does nothing.
What is the biggest weakness?
A summary model with a smaller context window than the main model. The docs say the middle of the conversation is then dropped without a summary.
Why does micro-compaction score lowest?
Each pass rewrites history the provider has already cached, so the prompt cache breaks every turn. It ships off by default.
📺 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
Rated 8 out of 10 overall, and if you want my team to check your setup against it, book a free SEO strategy session before you settle on Hermes compaction.
