openclaw - 💡(How to fix) Fix [Feedback] Dreaming produces irrelevant session summaries — material selection driven by word frequency rather than semantic importance [3 comments, 2 participants]

Issue Description

The /dreaming feature generates session summaries that are largely disconnected from the actual meaningful events discussed in our conversations. After running for ~10 days, the generated dreams feel like empty existential exercises rather than genuine reflections of our work together.

Concrete Example

Our real conversations on 2026-04-21 included:

• Buying advice for P5R (Persona 5 Royal) on Switch Hong Kong eShop
• Game comparisons: DQ VII remake vs P5R vs Super Robot Wars Y
• Nostalgia for a 2001 GAMEONE game 「永恒传说：奇幻水晶缘」
• Confirming the main characters were Kasper and Seber
• Discussing the game's “Will Seed” collectible mechanic

The actual dreaming output for that day was about counting “assistant” occurrences across 1400+ memories — completely ignoring the actual topics.

Root Cause Analysis

After reviewing the corpus files and session transcripts, the problem appears to be in the material selection layer, not the generation layer:

1. Corpus writes raw session transcripts including system metadata (Bootstrap warnings, HEARTBEAT_OK, JSON structures) — this creates noise
2. Material selection uses word frequency as the primary signal, not semantic importance
3. “assistant” appears ~1400+ times across sessions, so it always dominates the frequency stats
4. Meaningful decisions (like game purchases, tech configurations) appear once and are drowned out

Our Configuration

{
  "plugins": {
    "memory-core": {
      "config": {
        "dreaming": { "enabled": true }
      }
    },
    "active-memory": {
      "enabled": true,
      "config": {
        "agents": ["main", "wife"],
        "allowedChatTypes": ["direct"],
        "modelFallback": "minimax-portal/MiniMax-M2.7-highspeed",
        "queryMode": "recent",
        "promptStyle": "balanced",
        "timeoutMs": 15000,
        "maxSummaryChars": 220,
        "persistTranscripts": false,
        "logging": true
      }
    }
  }
}

This is a standard configuration. Nothing unusual.

Expected vs Actual

Expected: Dream summaries surfacing user decisions, conclusions, preferences, technical solutions found, non-obvious patterns noticed.

Actual: Generic existential reflections about the word “assistant” appearing N times.

Suggestion

Add a semantic importance filter in the material selection phase:

• Prioritize sessions with user decisions, conclusions, or acknowledgments
• De-prioritize sessions consisting primarily of single-word queries or system replies
• Consider adding a dreaming.promptStyle config to control material selection strategy

This is a product experience issue, not a bug per se. The feature works as designed — the design just doesn’t align with the expected utility.