Bug type

Behavior bug (incorrect output/state without crash)

Beta release blocker

No

Summary

The memory-core plugin's dreaming subsystem writes session transcripts to memory/.dreams/session-corpus/ — a directory inside the memory/ tree. The memory search indexer scans the entire memory/ directory, causing these session-corpus files to be indexed as regular memories. Over time this accounts for ~83% of recall store entries (1345/1623), and their content — raw JSON metadata, REM phase markers (<!-- openclaw:dreaming:rem:end -->), confidence: / evidence: patterns — gets promoted into MEMORY.md, polluting it with unreadable noise.

Steps to reproduce

Let the default dreaming cron (0 5 * * *) run for several days Observe memory/.dreams/session-corpus/ accumulating .txt session transcript files Run openclaw memory status — recall store shows hundreds of session-corpus entries Inspect MEMORY.md — entries with <!-- openclaw:dreaming:rem:end -->, confidence:, evidence: memory/.dreams/session-corpus/... markers appear at the bottom Check vector index size (grows disproportionately to useful memory content)

Expected behavior

Session-corpus files should not be indexed as memories MEMORY.md should only contain meaningful semantic memories, not REM reflection metadata short-term-recall.json entries should be < 300, not 1600+

Actual behavior

Session-corpus files at memory/.dreams/session-corpus/ are indexed by the memory search backend short-term-recall.json contains ~1345 session-corpus entries out of 1623 total MEMORY.md gets promoted entries containing raw REM metadata: Daily log files (memory/YYYY-MM-DD.md) are also polluted with <!-- openclaw:dreaming:rem:start --> / <!-- openclaw:dreaming:rem:end --> reflection blocks

OpenClaw version

2026.5.3-1 (commit 2eae30e)

Operating system

Ubuntu / cloud desktop (Linux 5.15.0)

Install method

npm global (openclaw update)

Model

MiniMax-M2.7-highspeed (primary), MiniMax-M2.7 (fallback)

Provider / routing chain

active-memory → MiniMax-M2.7-highspeed (blocking sub-agent) memory-search → DashScope text-embedding-v4 (OpenAI-compatible endpoint)

Additional provider/model setup details

"agents.defaults.memorySearch": { "enabled": true, "provider": "openai", "model": "text-embedding-v4", "remote": { "baseUrl": "https://dashscope.aliyuncs.com/compatible-mode/v1" } } "plugins.entries.memory-core.config.dreaming": { "enabled": true, "frequency": "0 5 * * *", "timezone": "UTC", "phases": { "light": { "lookbackDays": 3, "limit": 50, "dedupeSimilarity": 0.85 }, "rem": { "lookbackDays": 7, "limit": 30, "minPatternStrength": 0.5 }, "deep": { "limit": 20, "minScore": 0.3, "minRecallCount": 2 } } }

Logs, screenshots, and evidence

Recall store contamination (before cleanup):
Recall store: 1623 entries · session-corpus: 1345 (83%)
Vector store: 487 chunks (bloated)

Example corrupted entry in MEMORY.md:
## Promoted From Short-Term Memory (2026-04-16)
<!-- openclaw-memory-promotion:memory:memory/2026-04-15.md:196:233 -->
- <!-- openclaw:dreaming:rem:end --> --- # 2026-04-15 Daily Log...
  [score=0.594 recalls=4 avg=0.376 source=memory/2026-04-15.md:196-233]

Root cause locations in source code:
dreaming-phases-DW9aQqXD.js:796 — hardcoded path memory/.dreams/session-corpus
short-term-promotion-DZVrVqhT.js:397 — SHORT_TERM_SESSION_CORPUS_RE explicitly includes this path
short-term-promotion-DZVrVqhT.js:839 — isShortTermMemoryPath() returns true for session-corpus files

Impact and severity

Impact Detail Memory quality 83% of recall store is noise; useful memories buried MEMORY.md Polluted with REM metadata, unreadable by humans Active-memory Retrieval surfaces irrelevant session metadata Vector store Wastes embedding quota and storage on noise Severity Medium — does not break functionality but degrades all memory-dependent features

Additional information

Current workaround (manual, must be repeated after each dreaming run): Remove session-corpus entries from memory/.dreams/short-term-recall.json Strip <!-- openclaw:dreaming:(light|rem):start/end --> blocks from memory/YYYY-MM-DD.md daily files Run openclaw memory index --force to rebuild vector index Suggested fixes (in priority order): Change session-corpus output path to outside the memory/ tree (e.g., .openclaw/dreams/session-corpus/). This is the cleanest fix and fully prevents the issue. Add memorySearch.extraExcludePaths config option so users can exclude memory/.dreams/session-corpus from indexing. Add path filtering in short-term-promotion to skip session-corpus sources even when isShortTermMemoryPath() currently returns true for them.