[good first issue]🎯 [Bug] Hardcoded Chinese Headers in Persona/Scene Extraction Prompts Cause Language Leakage

[SD-fan]

OpenClaw Version | OpenClaw 版本

Hermes v0.15.1

Plugin Version | 插件版本

v0.3.6

Operating System | 操作系统

Ubuntu / Docker

System Specification | 系统配置

RTX Pro 6000, Qwen 3.6 27B FP8 MTP

Describe the bug | 问题描述

Issue: Hardcoded Chinese Headers in Persona/Scene Extraction Prompts Cause Language Leakage

Problem

The TencentDB Memory Gateway injects Chinese-language section headers into the Persona profile and memory context, even when the user communicates exclusively in German (or any other non-Chinese language). This happens because the prompt templates in scene-extraction.ts contain hardcoded Chinese section headers that the LLM follows as a template.

Affected Files

• src/core/prompts/scene-extraction.ts (lines 198-229)
• Generated persona.md files contain Chinese headers like:
  • ## 演变轨迹 (Evolution Trajectory)
  • ## 待确认/矛盾点 (Pending Confirmation / Contradictions)
  • ## 用户基础信息 (User Basic Information)
  • ## 用户核心特征 (User Core Traits)
  • ## 用户偏好 (User Preferences)
  • ## 隐性信号 (Implicit Signals)
  • ## 核心叙事 (Core Narrative)

Impact

1. Language contamination: Chinese characters appear in the injected memory context for every session, regardless of the user's language preference.
2. Agent confusion: The agent may accidentally copy Chinese headers into responses if it does not carefully filter the injected context.
3. User experience: Users who do not read Chinese see unexpected characters in their conversation context.
4. Persona profile pollution: The generated persona.md file contains Chinese section headers that persist across sessions.

Root Cause

The prompt templates in scene-extraction.ts use hardcoded Chinese section headers as output format instructions. The LLM follows these templates literally, producing Chinese headers in the generated persona profile. The CHANGELOG mentions that prompts should adapt to the user's language, but this is not implemented for the scene extraction templates.

Suggested Fix

Replace hardcoded Chinese headers in scene-extraction.ts with English or language-adaptive equivalents:

## 演变轨迹 -> ## Evolution Trajectory
## 待确认/矛盾点 -> ## Pending Confirmation / Contradictions
## 用户基础信息 -> ## User Basic Information
## 用户核心特征 -> ## User Core Traits
## 用户偏好 -> ## User Preferences
## 隐性信号 -> ## Implicit Signals
## 核心叙事 -> ## Core Narrative

Environment

• Hermes Agent: v0.15.1 (2026.5.29)
• TencentDB Memory Gateway: v0.3.6
• Model: qwen3.6-27b (custom provider)
• OS: WSL (Windows Subsystem for Linux)
• Docker: hermes-memory image

To Reproduce | 复现步骤

1. Set up TencentDB Memory Gateway with any LLM backend
2. Run conversations in German (or any non-Chinese language)
3. Check the generated persona.md file
4. Observe Chinese section headers in the output

Expected behavior | 预期行为

The prompt templates should either:

• Use language-neutral headers (English as default)
• Dynamically adapt headers to the detected user language
• At minimum not inject Chinese characters into non-Chinese conversations

Error Logs / Screenshots | 报错日志/截图

No response

Additional context | 补充信息

No response

[Maxwell-Code07]

Thank you for your detailed report. Regarding Chinese chapter titles and language type adaptation, we have indeed paid attention to it internally before. We will evaluate this demand internally as soon as possible. Thank you for your clear explanation!

[Maxwell-Code07]

We have selected this Issue as a 🎯 Good First Issue — perfect for new contributors to get started!

Task Details

Field	Content
Type	Bug
Difficulty	Medium
Estimated Effort	5 days
Acceptance Criteria	1. Replace hardcoded Chinese headers in prompts with dialogue language adaptation; 2. Memory titles output in English during English conversations; 3. Add corresponding unit tests

Feel free to ask questions if you have any! 🚀

[OS-Lihua]

claim

[sirenexcelsior]

I have joined the TencentDB Agent Memory Open Source Co-creation Program and claimed this issue.