Skip to content

docs: clarify how media attachments affect session context #25650

@evgyur

Description

@evgyur

Problem

Users can reasonably misunderstand how Hermes' session context treats media attachments. The current Sessions documentation explains persistence and session storage, but it does not spell out the distinction between:

  • raw media/file bytes
  • paths or metadata saved in the transcript
  • extracted text, transcriptions, or vision summaries
  • verbose tool outputs/logs that actually grow the active context

That gap makes people worry that sending one image or file permanently bloats every later model call in the same session.

Why this matters

Hermes is frequently used through gateway platforms such as Telegram, Discord, and Slack, where users attach images, voice messages, PDFs, text files, logs, and screenshots. They need a simple mental model for context cost and privacy:

  • media is usually a turn-scoped input;
  • only text/metadata/derived analysis that Hermes injects into the conversation becomes regular prompt history;
  • the real context-growth hazard is dumping long logs, transcripts, diffs, and proof reports into chat.

Clear docs reduce support/debug churn and help users choose /compress, /new, summaries, file paths, or focused excerpts correctly.

Example from today

A user sent an image and asked Hermes to make a meme from it. Later in the same Telegram topic, they asked whether that image/meme would keep expanding the model context in future turns.

The right explanation was:

  • the original JPEG is not re-sent forever just because it was attached once;
  • Hermes may analyze the image once with vision or attach it natively for one model call;
  • future turns only carry what was actually written into the transcript, such as the user's request, a local cache path, a short image description, or the assistant's response;
  • the bigger context risk is verbose text: pasted transcripts, full logs, long diffs, repeated status updates, and proof dumps.

That explanation belongs in the Sessions docs.

Proposed change

Add a short section to website/docs/user-guide/sessions.md explaining what counts toward context, how media attachments are handled, and when to use /compress, /new, or session pruning.

Metadata

Metadata

Assignees

No one assigned

    Labels

    P3Low — cosmetic, nice to havetype/docsDocumentation improvements

    Type

    No type
    No fields configured for issues without a type.

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions