Skip to content
vavio

Does changing Hermes's model keep the same conversation?

Yes: switching Hermes Agent’s model within an existing conversation keeps its stored history. Hermes changes the model connection and prepares subsequent requests for the new route. That does not guarantee the new model sees every original detail, supports the same tools, or accepts the same request.

This answer follows Hermes v0.23.1, commit 3b42157, checked on 13 November 2026. We reviewed official source and documentation; we did not run a model-switch trial on a Vavio computer. The behavior below concerns Hermes’s own model-switch paths, not a promise that editing a configuration file changes an already running agent.

What stays, and what changes?

Part of your work What a model switch means
Stored conversation history The switch does not clear the transcript. With default in-place compression, later compaction keeps the same session ID and archives older turns under it; the older in_place: false setting creates a continuation session instead.
Context sent with the next request Hermes prepares the current conversation window for the new route. Earlier details may already have been summarized, and original attachment bytes are not automatically resent forever.
Model connection Hermes changes the model, endpoint and API mode as needed, rebuilds the client, and refreshes provider-related runtime settings.
Context limit and compression Hermes updates the compressor’s model and context limit. A smaller window can require compression before the next reply.
Persistent memory Switching does not erase the profile’s memory files. Those files are separate from both the transcript and the current prompt.
Tools and capabilities The destination model and provider must support the requests you need. An ordinary text reply does not prove tool use works.

In the terminal CLI, Hermes switches the existing agent in place and preserves the conversation. The messaging gateway first switches a cached agent if present, records the new session route, then evicts that agent so the next turn builds a fresh one from the override. Both paths preserve transcript history; they reach the next reply through different runtime steps. See the linked official source above for each path.

Will a cheaper or smaller model remember everything?

It can continue from the context Hermes supplies, but it may receive a summary instead of every original turn. A smaller context window can force compression on the next message. The CLI has a warning for this case, but absence of a warning does not prove that compression cannot happen.

Treat exact requirements as saved work: keep the task brief, decisions, file paths and acceptance checks in a document you can read back. If a detail matters to the result, ask the new model to identify it from that document or the available conversation before relying on it. A fluent continuation alone does not show that every detail survived.

A larger model window also does not recreate details already omitted from the current context. Hermes can search stored sessions when asked; that is retrieval, not automatic inclusion of every past message. What Hermes remembers after a new chat explains the separate roles of history and persistent memory.

Does changing provider make a difference?

Yes. Providers can expect different message formats, reasoning fields, tool-call identifiers and image inputs. Hermes includes request adaptations, including repair of tool-call/result pairing. Those adaptations do not establish universal compatibility: a destination can still reject a request or lack a capability your task uses.

If the next request fails, distinguish a connection or authentication error, an unsupported context window, and a tool-calling failure before treating it as lost memory. Why Hermes rejects a model or ignores tools gives the checks for the latter two cases. This article does not rank model quality or compare their task performance.

Does my saved memory move to the new model?

Your profile’s memory files remain on the computer; they do not become knowledge trained into the destination model. Hermes supplies memory through its own context machinery. Its documented session-start memory snapshot also means you should not assume that changing the model immediately reloads a memory file you just edited.

Changing model within a conversation, starting a new chat, and choosing a different profile are different actions. A profile has its own state and model configuration. How two profiles can share a model account explains those boundaries.

How should I continue an important task after switching?

Use this as a continuity check, not a claim that we ran the sequence on every provider:

  1. Save a short checkpoint: desired output, completed steps, remaining work, important constraints and the files that contain the evidence. Let any current tool action finish before changing routes.
  2. Use the model-switch path for your Hermes interface and verify the selected model/provider. Changing a default setting alone is not evidence that the current conversation now uses it.
  3. Ask the new model to restate the remaining task and identify the saved checkpoint. Correct any missing constraint before asking it to continue.
  4. Try one small action needed by the task, then inspect the actual file or tool result. A text claim that the action worked is insufficient.

If you want Hermes available after your laptop closes, Vavio provides its own managed computer for $14.99 per month, with model usage billed separately by your provider. Hosting keeps the computer available; choosing a model that can finish your particular job still needs a task check. Compare the current hosting offer when deciding between managed hosting and running Hermes yourself.

Sources

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 model-switch trial.

More from the Vavio blog: Hermes Agent guides