Summary

During overnight dreaming, the light-phase work summary block emits identical commit-list content every cycle for ~6 hours, producing apparent duplicates in DREAMS.md. Only the REM reflection block (which precedes each work summary) shows incrementing memory counts, suggesting the REM input pipeline is incremental but the work-summary pipeline is not.

Evidence (inline excerpts)

DREAMS.md is auto-published from ~/.openclaw/memory/DREAMS.md. Two adjacent cycles 29 minutes apart, on the same workspace:

April 12, 2026 at 2:00 AM (cycle 1):

Reflections: Theme: `assistant` kept surfacing across 230 memories.;
  confidence: 0.96; evidence: memory/.dreams/session-corpus/2026-04-10.txt:2-2,
  memory/.dreams/session-corpus/2026-04-10.txt:4-4, ...; note: reflection

- Added primary issue extraction for pain notifications so notify
  output explicitly states the shared main issue instead of only
  saying "多源共振". - Updated signals cron notification style ...
  commits: `1124b18` `fix: clarify signals pain notifications` ...

April 12, 2026 at 2:29 AM (cycle 2, 29 min later):

Reflections: Theme: `assistant` kept surfacing across 397 memories.;
  confidence: 1.00; evidence: memory/.dreams/session-corpus/2026-04-10.txt:2-2,
  memory/.dreams/session-corpus/2026-04-10.txt:4-4, ...; note: reflection

- Added primary issue extraction for pain notifications so notify
  output explicitly states the shared main issue instead of only
  saying "多源共振". - Updated signals cron notification style ...
  commits: `1124b18` `fix: clarify signals pain notifications` ...

The REM reflection block went 230 memories → 397 memories (so the cycle did fire fresh and saw new corpus), but the work-summary block is byte-for-byte identical, including the same commit hash 1124b18.

This pattern continues across the same night for 6 cycles (2:00 / 2:29 / 3:29 / 4:29 / 5:29 / 6:29 AM) and the next night for another 5 cycles, producing the same 1124b18 work-summary block emitted verbatim 11 times across two nights. REM memory count climbs through 230 → 397 → 551 → 705 → 846 → 987 → 1239 → 1338 → 1430 → 1526 → 1604 across these cycles.

The full DREAMS.md is in a private repo; happy to share full samples or a session bundle on request.

Steps to reproduce

1. Have an active workspace with git commits in the last ~24h
2. Let dreaming run overnight (heartbeat decoupling per memory/dreaming: decouple managed cron from heartbeat #70737 active)
3. Observe DREAMS.md the next morning — work-summary content is identical across cycles

Expected behavior

Light-phase work summary should be incremental: only commits that are new since the last cycle's emit should appear. Alternatively, content could be dedup'd against prior entries within the same dreaming session window.

Actual behavior

Each cycle re-summarizes the same fixed window (apparently "last ~24h commits"), without tracking a "last-emitted cutoff" or any dedup against the prior cycle's output. Result: the same commit-list block repeats across every cycle until the underlying commit window rolls forward.

Environment

• OpenClaw 2026.4.24
• Heartbeat decoupling (memory/dreaming: decouple managed cron from heartbeat #70737) active
• Config: minimal — enabled: true, frequency: "0 3 * * *", storage.mode: "separate". No phase-level overrides.
• Channel: managed dreaming cron + light-phase implicit schedule

Suspected mechanism

Light/deep cycle's commit-summarizer reads a fixed time window from git history (or a corpus snapshot) without tracking a per-phase cutoff timestamp, so identical input → identical output every cycle.

Proposed fix sketch

Persist a per-phase lastEmittedAt in dreaming state; summarizer filters input to commits.timestamp > lastEmittedAt before generating the summary, then advances lastEmittedAt on successful emit.

Related

• Deep dreaming promotes zero candidates: recallCount stays at 0 and maxScore caps at 0.62 #68882 (closed) added isContaminatedDreamingSnippet filter for promotion candidacy — different layer; that filter does not affect diary write path, which is what this bug is about.
• memory/dreaming: decouple managed cron from heartbeat #70737 dreaming/heartbeat decoupling — orthogonal; happens upstream of this symptom.