Summary

The memory search vector index stops updating after the chokidar file watcher silently stops firing events. Files modified after this point are never re-indexed, causing memory_search to return 0 results for content that exists on disk. No error or warning is logged when the watcher fails.

Environment

• OpenClaw: 2026.3.2 (85377a2)
• OS: Ubuntu 24.04, kernel 6.17.0-14-generic (x86_64)
• Node: v22.22.1
• Memory search provider: openai / text-embedding-3-small
• Config: default (no custom memorySearch settings)

Reproduction

1. Start gateway, let it run for several hours with active agent sessions
2. Agent tools (exec/write) modify files under memory/ and memory/subdir/ (e.g. daily notes, reference docs)
3. Run memory_search for content that exists in recently modified files
4. Expected: Results from the modified files
5. Actual: 0 results. The SQLite index (~/.openclaw/memory/main.sqlite) shows stale size/mtime for the modified files

Investigation

Watcher setup (correct):
The code at manager-*.js line ~2244 creates:

chokidar.watch([
  "<workspace>/MEMORY.md",
  "<workspace>/memory.md",
  "<workspace>/memory/**/*.md"
], { ignoreInitial: true, awaitWriteFinish: { stabilityThreshold: 1500, pollInterval: 100 } })

inotify watches (incomplete):
Inspecting /proc/<gateway-pid>/fdinfo/<inotify-fd> shows only 8 watches. The memory/ directory has a watch, but subdirectories under it (e.g. memory/subdir/) do not — despite the **/*.md glob which should trigger recursive watching via chokidar.

Timeline:

• Gateway started at T+0
• Last successful [memory] sync: indexing memory files log entry at T+1h (triggered by a session that modified files)
• After T+1h, 15+ files were modified by various agent sessions over the next 20 hours
• Zero sync events occurred despite file changes
• Flat memory/*.md files also not synced (not just subdirectory files)

No fallback:
sync.intervalMinutes defaults to 0 (disabled), so when the watcher stops working there is no periodic fallback to catch stale files.

Workaround

Setting agents.defaults.memorySearch.sync.intervalMinutes: 30 provides a safety net. openclaw memory index --force can manually trigger a full reindex.

Suggested fix

1. Add a periodic health check / heartbeat for the watcher (e.g. verify it's still receiving test events)
2. Consider defaulting intervalMinutes to a non-zero value (e.g. 15-30 min) as a fallback
3. Log a warning if the watcher's ready event doesn't fire or if no change events are received for an extended period despite files being modified