Repository navigation
Description
Bug type
Regression (worked before, now fails)
Beta release blocker
No
Summary
After upgrading from 2026.5.26 to 2026.6.1 with memorySearch.provider: gemini, memory status reports inconsistent Dirty state between bare and --deep invocations, and the memory watcher no longer auto-indexes new files.
Steps to reproduce
Before upgrade, set openclaw config set agents.defaults.memorySearch.provider gemini
Upgrade 2026.5.26 → 2026.6.1
Run openclaw memory status → shows Dirty: yes, Vector search: paused
Run openclaw memory status --deep → shows Dirty: no, Vector store: ready
Write new files to memory extra path (~/clawd/memory/)
Wait — watcher does not trigger, files remain unindexed
Run openclaw memory status --deep → still shows previous indexed count
Run openclaw memory status --index --agent main → successfully indexes new files
Expected behavior
memory status and memory status --deep should report consistent health state
Memory watcher should auto-detect and index new files as in 2026.5.26
Manual --index should not be required for routine file additions
Actual behavior
memory status reports Dirty: yes / vector search paused (misleading)
memory status --deep reports healthy state (correct)
Watcher silent — gateway.log shows no memory/embed/sync activity after startup
New files only indexed after manual openclaw memory status --index --agent main
OpenClaw version
2026.6.1 (2e08f0f)
Operating system
Rocky Linux, Node v24.14.1
Install method
npm global (openclaw update)
Model
google/gemini-3-flash-preview
Provider / routing chain
Google AI Studio API key → OpenClaw Gateway → memorySearch (gemini/gemini-embedding-001)
Additional provider/model setup details
json"memorySearch": {
"extraPaths": ["/path/to/memory/"],
"enabled": true,
"provider": "gemini"
}
Logs, screenshots, and evidence
gateway.log after startup — no memory watcher activity:
2026-06-06T08:38:43.230+08:00 [gateway] http server listening (9 plugins: memory-core, telegram, ...)
2026-06-06T08:41:05.221+08:00 [gmail-watcher] gmail watcher stopped
(No memory/embed/sync/watcher entries)
memory status (misleading):
Provider: gemini (requested: gemini)
Dirty: yes
Vector search: paused until memory is rebuilt
Index identity: index was built for model gemini-embedding-001, expected
memory status --deep (correct):
Provider: gemini (requested: gemini)
Model: gemini-embedding-001
Dirty: no
Vector store: ready
Semantic vectors: ready
Embeddings: ready
After manual --index:
Memory index complete.
Indexed: 1384/1384 files · 21580 chunks
Dirty: noImpact and severity
Affected: Users with Google/gemini memorySearch provider on 2026.6.1
Severity: Medium — memory health status is misleading, auto-indexing broken. Workaround available (manual --index), but automation pipelines relying on watcher will silently fail.
Frequency: Always reproducible after upgrade
Additional information
Related to #90786 and #90787 (same upgrade path, same provider configuration).
Workaround for #90786/#90787: setting memorySearch.provider: gemini before upgrading preserves vector search functionality. However this issue (#watcher + Dirty display) persists even with the workaround applied.
Last known good: 2026.5.26 — watcher auto-indexed new files immediately, memory status and --deep reported consistent state.
Metadata
Metadata
Assignees
Labels
Type
Fields
Priority
Activity
clawsweeper commentedon Jun 6, 2026
Thanks for the report. I checked current
main, but this shell is missing enough reproduction detail to verify a current bug.Close as stale: the only human follow-up corrected the watcher claim to delayed-but-eventual indexing, and no current-version reproduction has appeared in over 60 days while the status and Linux watcher paths have changed substantially.
Review details
Best possible solution:
Keep the centralized watcher and identity paths; reopen with a current-release reproduction if a timestamped extra-path write still takes hours to index.
Do we have a high-confidence way to reproduce the issue?
No. The supplied evidence is limited to v2026.6.1, and the reporter later observed automatic indexing resume; there is no current-release reproduction.
Is this the best way to solve the issue?
Yes. A stale close is safer than a speculative repair because current source covers the watcher and status paths, while the remaining provider reports are separately tracked.
Security review:
Security review: This issue does not report a security-sensitive behavior.
AGENTS.md: found and applied where relevant.
What I checked:
extensions/memory-core/src/memory/manager-watch-ops.ts:127, d6233c29b904)extensions/memory-core/src/memory/manager.watcher-config.test.ts:1002, d6233c29b904)extensions/memory-core/src/memory/manager-sync-base.ts:349, d6233c29b904)extensions/memory-core/src/memory/manager-watch-ops.ts:399, 5a67c5c5566e)Likely related people:
extensions/memory-core/src/memory/manager-watch-ops.ts)extensions/memory-core/src/memory/manager-watch-ops.ts)extensions/memory-core/src/memory/manager-sync-ops.ts)Codex review notes: model internal, reasoning high; reviewed against d6233c29b904.
fenglanhua commentedon Jun 6, 2026
Update after further observation:
After monitoring throughout the day, main.sqlite did update automatically — Indexed count increased from 1384/1384 (21580 chunks) after manual --index at 08:59, to 1387/1387 (21589 chunks) at 14:24, without any manual intervention.
This suggests the watcher is not completely broken, but has significant latency — new files took several hours to be automatically embedded, compared to near-instant behavior in 2026.5.26.
Noting that 2026.6.1 changelog mentions reduce Linux watcher fan-out as an intentional change. If this latency is by design (e.g. to reduce CPU load on low-power hardware), the issue may be better framed as: the new watcher behavior needs clearer documentation, and operators relying on timely memory indexing should be advised to use openclaw memory status --index manually or via scheduled task.
Happy to provide more data if useful for triaging whether this is intentional or a regression.
5 remaining items
vincentkoc commentedon Aug 12, 2026
Closing as stale. The false
Dirty: yesstatus was fixed by #90816 and shipped in v2026.7.1-2. The reporter later corrected the total watcher-failure claim: the index advanced automatically, leaving only old, unquantified latency without a current Rocky Linux reproduction. Broader watcher reliability remains tracked at #40088 and #119411. Please open a fresh report with timestamped evidence from the current release if this still reproduces.