Why did Hermes compress my conversation, and can I get the original detail back?
Hermes Agent compresses a conversation when it grows close to the model’s context limit. It replaces older turns with a summary so the next request fits. With default settings the original turns are not deleted: they stay in the session’s stored history, and Hermes can search them to recover an exact detail the summary left out.
This answer follows Hermes v0.23.1, commit 3b42157, checked on 14 November 2026. We reviewed official source and documentation; we did not run a compression or recovery trial on a Vavio computer. Settings below are Hermes defaults. A profile that changed them, or uses a context-engine plugin, can behave differently.
When does Hermes compress?
Hermes’s own compressor runs inside the conversation when the context it would send passes a threshold. The configured default is 50% of the model’s usable window, but models with windows under 512K tokens are raised to at least 75%. So a model with a window under 512K tokens compresses at about three-quarters of it; a model with a 512K or larger window compresses at half. compression.threshold_tokens sets an absolute trigger, and Hermes uses whichever is lower.
Messaging conversations have a second safety net. Before a new message reaches the agent, the gateway checks whether the session has passed 85% of the window, for example after a Telegram chat has accumulated overnight. Either path ends in a summary of the older turns.
The Hermes sessions guide says verbose text, such as pasted transcripts, full logs, large tool output and long diffs, is the most common cause of context growth, not attached media.
Will I see it happen?
It depends on where you talk to Hermes. On messaging platforms such as Telegram or Discord, routine automatic compression is silent by default. compression.progress_notices: true turns on start and completion notices. Failure notices and the reply to a manual /compress always appear.
The usual sign is a reply that seems to have lost an earlier specific: an exact error message, a number you gave it, or a constraint stated early in the chat.
What does the summary keep?
The summary is written by a separate model call. It uses your main model unless auxiliary.compression.model names another. In the default “lean” mode it contains:
| Part | What it carries |
|---|---|
| Narrative summary | Goal, constraints, completed actions, current state, decisions and relevant files. |
| Your messages | Your real messages from the compacted part, quoted word for word, newest first, up to about 24,000 characters (each long message trimmed to 4,000). |
| Session log | A dense chronological log the summarizer is told to write with exact identifiers: file paths, commands, error text, version numbers. |
| Anchor index | Issue numbers, commit hashes, file paths, error names and URLs extracted mechanically, without a model. |
| Recovery note | The number of compacted messages and a session_search pointer with this session’s ID. |
The most recent part of the conversation stays verbatim alongside the summary, as do the system prompt and the first few messages of the conversation. Old tool output in that recent part can be replaced by a one-line note saying how many characters were moved to session history.
What the summary can lose is anything the model writing it judged unimportant: a long command output, a file Hermes read, earlier reasoning or wording it paraphrased. A cheaper summary model writes a weaker summary. Hermes marks the summary as reference material and tells the model to answer only the latest message, so a request already handled before compression should not be repeated.
Can I get the original detail back?
Yes, if your profile uses the default in-place compression (compression.in_place: true). Hermes keeps the same session ID and marks the older turns as compacted instead of deleting them. Its session_search tool searches those compacted turns as well as live ones and returns the stored messages themselves, not a new summary. Very long messages come back clipped, and Hermes can page around a message for more. By default the search covers your messages and Hermes’s replies; tool output, such as a command’s result, is searched only when Hermes asks for it.
The simplest recovery is to ask in the same conversation, for example: “Before the context was compressed I gave you the exact error from the deploy. Search this session’s history for it and quote it.” The summary’s recovery note already gives Hermes this session’s ID. For an exact command output, add “include tool output”. If Hermes answers from memory instead, ask it to use session_search and quote the message.
To keep a readable copy of the full conversation, /save md exports the session’s history as Markdown, including turns archived by in-place compression. On a messaging platform Hermes sends the file into the chat as a document; in the terminal CLI it writes the file on the machine running the CLI. With a terminal on the Hermes computer, hermes sessions export --format md --session-id <id> does the same. Add redact (for example /save md redact) to scrub keys and tokens before you share the file. These are exports, not undo: they do not put the old turns back into the model’s working context.
With the older setting in_place: false, compression starts a new continuation session (titled “my project #2” and so on) instead. The earlier session still holds the original turns, so session search and export still find them.
When is the detail actually gone?
Compression does not delete stored messages, and it is not a privacy delete either. The original detail is gone from storage only if the session is deleted or pruned. Hermes’s automatic pruning of old ended sessions is off by default.
If compression keeps failing and the request no longer fits the model, Hermes first falls back to a fixed summary written without a model call. As a last resort it ends the turn, and on messaging platforms it may start a new session. Treat that as a reason to start a fresh chat deliberately rather than to keep extending one thread.
How do I avoid losing important detail?
- Put requirements you must not lose in a file: the task brief, constraints, acceptance checks and paths to the evidence. Hermes can reread a file; a summary may paraphrase a chat message. Which instructions Hermes loads for a project explains where standing instructions belong.
- Ask Hermes to keep long output in files and report a short excerpt, rather than pasting full logs into chat.
- Compress on your terms.
/compress here 4keeps your last four exchanges verbatim and summarizes the rest./compress <topic>tells the summary what to preserve. - Start a new chat with
/newwhen you finish a task. The Hermes sessions guide notes that memory files are loaded at session start, so a long-running thread uses saved memory less. What Hermes remembers after a new chat covers that boundary. - If you switch to a model with a smaller window, expect compression sooner. Does changing Hermes’s model keep the conversation? explains that case.
Long-running conversations and overnight messaging threads are exactly where compression happens, and they are also why people keep Hermes on a computer that stays online. Vavio gives your Hermes its own managed computer for $14.99 per month, with model usage billed separately by your provider. Hosting keeps the conversation reachable; it does not change how Hermes compresses it. Compare the current hosting offer with running Hermes yourself.
Sources
- Hermes v0.23.1 context compression settings, checked 14 November 2026
- Compression defaults and small-window floor, checked 14 November 2026
- Summary contents and recovery pointer, checked 14 November 2026
- Session search includes compacted turns, checked 14 November 2026
- Hermes v0.23.1 sessions guide, checked 14 November 2026
- Vavio hosting offer, checked 14 November 2026
Written by Vavio with AI assistance and checked against pinned official Hermes v0.23.1 source and documentation. This is a source review, not a live compression or recovery trial.