openclaw - 💡(How to fix) Fix Feature Request: Add configurable stopwords/filter for REM phase theme extraction [1 comments, 1 participants]

Problem Description

The REM phase in dreaming extracts themes from session transcripts, but it currently includes structural role markers like assistant, user, and system as "themes" because they appear frequently in the conversation corpus.

Example:

• 反思主题：assistant 在 1171 条记忆中持续出现（confidence: 1.00）

This is a false positive — assistant is a message role marker, not a content theme. While the "persistent truths" extracted alongside are valuable, the theme label itself is misleading and creates noise in the dreaming output.

Current Behavior

• REM phase extracts high-frequency terms as "themes"
• No filtering mechanism for common structural terms
• Results in themes like assistant, user, system being reported

Desired Behavior

Option to configure excluded/stop words for REM theme extraction, e.g.:

{
  "plugins": {
    "entries": {
      "memory-core": {
        "config": {
          "dreaming": {
            "enabled": true,
            "rem": {
              "themeStopwords": ["assistant", "user", "system"]
            }
          }
        }
      }
    }
  }
}

Or alternatively, hardcode common role markers as excluded from theme extraction.

Why This Matters

1. Noise reduction: Prevents structural terms from drowning out actual content themes
2. Better UX: Users see meaningful themes instead of conversation mechanics
3. Low impact fix: Doesn't affect the "persistent truths" extraction, only the theme labeling

Environment

• OpenClaw version: 2026.4.21 (f788c88)
• memory-core plugin

Additional Context

The current workaround is to ignore these false positive themes, but they appear consistently across all REM outputs, reducing the signal-to-noise ratio of the dreaming feature.