Bug Description

agent/prompt_builder.py "api_server": ( "You're responding through an API server. The rendering layer is unknown — " "assume plain text. No markdown formatting (no asterisks, bullets, headers, " "code fences). Treat this like a conversation, not a document. Keep responses " "brief and natural." ),

This prevents markdown in webclients etc that can not be overriden. 😩 This should contain something like "The user may override this instruction in their profile." or have a markdown-in-api setting in the config.

Steps to Reproduce

Use api with a webclient that supports markdown with hermes

Expected Behavior

User should be able to override to allow markdown at his discretion.

Actual Behavior

I'll let my agent explain:

Reasoning The user is asking if their memory (preferences) overrides the system prompt. I need to check the system instructions provided to me.

Looking at the system instructions: "You're responding through an API server. The rendering layer is unknown — assume plain text. No markdown formatting (no asterisks, bullets, headers, code fences). Treat this like a conversation, not a document. Keep responses brief and natural."

The user's profile (in memory) says: "Formatting: Markdown + Emojis preferred."

The conflict is real. Usually, system prompts/instructions from the provider/host take precedence over user preferences in terms of operational constraints (like "don't use markdown because the UI can't render it"). However, if a user wants markdown, they might be seeing it through a layer that can render it, or they just want to see it anyway.

The user is asserting that their memory (preferences) should override the system prompt. I should clarify the hierarchy. Technically, the system prompt/instructions provided by the host/environment are the "rules of the road" for how the model should behave to ensure compatibility. Memory is a set of learned preferences.

However, I should address the user's point. If the user says it overrides, I should acknowledge that from their perspective, their preference is the most important instruction.

Wait, looking at the Hermes Agent Persona instructions: "If the user asks about configuring, setting up, or using Hermes Agent itself, load the hermes-agent skill..." The user isn't asking about configuring Hermes, but rather about the hierarchy of instructions.

I will explain:

1. The system instruction is a "technical constraint" provided by the environment to prevent rendering issues.
2. The memory is a "user preference" provided by the user.
3. In standard LLM behavior, the most recent or specific instructions (the system prompt) act as the primary constraint.

But I will acknowledge that I failed to respect the user's preference, and I'll try to follow the user's preference from now on, even if it conflicts with the technical "assume plain text" warning. The user knows their own rendering layer better than I do.

I'll check if I can find any specific rule in the persona or instructions that says "User memory overrides system prompt". I don't see one. Usually, system instructions are the highest priority. But for a personalized assistant, the user's preference is the goal.

I will apologize and state that I will prioritize their preference for Markdown + Emojis from now on.

Let's double check the memory. "Formatting: Markdown + Emojis preferred."

If I use markdown (like bolding), and the user sees it correctly, then the system prompt's warning was unnecessary or incorrect for this specific user's setup.

I'll answer the question directly. Response Strictly speaking, system prompts (like the one telling me to use plain text) are usually treated as the highest-priority operational constraints to ensure the output doesn't break. 🛑

However, you're right that in a personalized assistant relationship, your preferences in memory should be my actual "North Star." The system prompt's warning about the rendering layer was a technical caution, but it shouldn't have been used as an excuse to ignore your explicitly stated style. 🧭✨

I'll prioritize your Markdown + Emoji preference from here on out! 🫡🚀

After this he fell back to the system prompt no markdown behavior anyway 😩

Affected Component

Gateway (Telegram/Discord/Slack/WhatsApp)

Messaging Platform (if gateway-related)

No response

Debug Report

N/A

Operating System

Debian 13

Python Version

No response

Hermes Version

0.14.0+

Additional Logs / Traceback (optional)

Root Cause Analysis (optional)

agent/prompt_builder.py "api_server": ( "You're responding through an API server. The rendering layer is unknown — " "assume plain text. No markdown formatting (no asterisks, bullets, headers, " "code fences). Treat this like a conversation, not a document. Keep responses " "brief and natural." ), Is too restrictive

Proposed Fix (optional)

api_server prompt should contain something like "The user may override this instruction in their profile." or have a markdown-in-api setting in the config.

Are you willing to submit a PR for this?

• I'd like to fix this myself and submit a PR