Bug type

Regression (worked before, now fails)

Beta release blocker

No

Summary

Light Sleep phase in Dreaming produces all candidates with confidence=0.00, preventing any meaningful memory promotion. REM Sleep phase works correctly with normal confidence values (~0.70).

Steps to reproduce

1. Enable Dreaming in plugins.entries.memory-core.config.dreaming.enabled: true
2. Let Dreaming run for a few cycles (every 30 min)
3. Check memory/YYYY-MM-DD.md for Light Sleep section
4. Observe all candidates have confidence: 0.00

Expected behavior

Light Sleep candidates should have non-zero confidence values based on semantic relevance, similar to REM Sleep candidates which show confidence ~0.70.

Actual behavior

In memory/2026-04-11.md, all Light Sleep candidates show confidence: 0.00 with status: staged. Example:

• Candidate: ...
  • confidence: 0.00
  • evidence: memory/2026-04-11.md:83-86
  • recalls: 0
  • status: staged

REM Sleep candidates work correctly with confidence ~0.70.

OpenClaw version

2026.4.9 (build 0512059)

Operating system

Windows 10 x64

Install method

npm global

Model

astroncodingplan/astron-code-latest

Provider / routing chain

openclaw -> astroncodingplan

Additional provider/model setup details

No response

Logs, screenshots, and evidence

Impact and severity

No response

Additional information

This appears to be a new bug - searched GitHub issues and found no matching report for Light Sleep confidence=0.00. Related issues found but different:

• Control UI: Dreaming page layout/rendering regression in 2026.4.9 #63721: Control UI rendering issue (closed)
• [Bug]: Dreaming/REM surfaces raw session-corpus metadata (untrusted, user, assistant) instead of meaningful reflections #63921: REM output noise issue
• fix(hooks): preserve plugin-registered internal hooks when refreshing file hooks #64340: Dreaming startup bug (fixed)

None address the Light Sleep confidence calculation bug.