openclaw - 💡(How to fix) Fix Dreaming: REM scorer promotes line-number references and filenames as themes [1 participants]

Summary

The REM sleep phase of the dreaming system is pattern-matching on citation metadata (line number ranges like 28-31, 13-16, and file paths like memory/2026-04-12.md) and promoting them as meaningful recurring themes.

Observed behavior

After the first successful dreaming run (2026-04-18), two agents (a coding agent and a math/tutoring agent) both had entries like these promoted into MEMORY.md:

Theme: `28-31` kept surfacing across 39 memories.
  - confidence: 1.00
  - evidence: memory/2026-04-12.md:43-46, ...
  - note: reflection

Theme: `13-16` kept surfacing across 41 memories.
  - confidence: 1.00
  - evidence: memory/2026-04-12.md:43-46, ...
  - note: reflection

Theme: `memory/2026-04-12.md` kept surfacing across 56 memories.
  - confidence: 1.00
  - ...

These strings are line citation ranges from the evidence field format (e.g. source=memory/2026-04-12.md:28-31), not actual conceptual themes from session content.

Root cause (suspected)

The REM scorer appears to tokenize or extract "themes" from the full evidence/source string rather than from the semantic content of the candidate text itself. When candidates reference the same source file repeatedly (as expected when dreaming re-processes its own prior dream outputs), the line range suffixes and filenames become the most frequently occurring tokens and get ranked as themes.

Impact

• Noise entries are permanently promoted into MEMORY.md with confidence 1.00
• Affects agents with lower session volume (e.g. a coding agent, a math/tutoring agent) more severely — their small session corpus means self-referential dreaming output dominates the token frequency
• Creates a compounding problem: once these are in MEMORY.md, they become part of the corpus for future dreaming sweeps

Expected behavior

The REM scorer should:

1. Filter out tokens that match line-range patterns (e.g. /^\d+-\d+$/) from theme extraction
2. Filter out file path tokens (e.g. strings containing / or ending in .md) from theme extraction
3. Ideally, extract themes from the candidate text content rather than from the citation/evidence metadata

Environment

• OpenClaw: v2026.4.12
• Dreaming config: minScore: 0.8, minRecallCount: 3, minUniqueQueries: 3
• Memory backend: QMD bridge mode, BM25-only (no vector embeddings — ARM host, no Vulkan)
• Agents affected most severely: low-volume agents (coding agent, math/tutoring agent); also seen in a research agent