Memory store unavailable — database wiped on gateway restart, startup checks never log

[DSchexy]

Bug Description

The memory-lancedb-pro plugin registers successfully but shows "unavailable" status. The LanceDB database appears to be wiped on every gateway restart, and the startup checks never produce any log output.

Environment

• OpenClaw version: 2026.3.23-2 (gateway managed as standalone process)
• Plugin version: 1.0.32 (npm global install) — also tested 1.1.0-beta.9
• LanceDB version: @lancedb/lancedb 1.x
• Embedding model: nomic-embed-text (768 dims, Ollama local)
• OS: macOS 25.3.0 (arm64)
• Node: v22.22.1
• Database path: /Users/luna/.openclaw/memory/lancedb-pro

Steps to Reproduce

1. Install memory-lancedb-pro globally: `npm install -g memory-lancedb-pro`
2. Configure in OpenClaw config with LanceDB path + Ollama embeddings
3. Start gateway — plugin registers ("memory-lancedb-pro@X.X.X: plugin registered" appears in logs)
4. Run `openclaw status` — Memory shows "unavailable"
5. Check `~/.openclaw/memory/lancedb-pro/` — directory exists but is empty (no memories.lance)
6. Manually create LanceDB tables — tables exist, data is present
7. Restart gateway — database is wiped again, status returns to "unavailable"

Expected Behavior

• Plugin should initialize LanceDB on first start without wiping existing data on restart
• Startup checks should log either "initialized successfully" or "startup checks failed" with error details
• Memory status should show "available" once tables are created

Actual Behavior

1. Empty database on first start: Plugin creates the directory structure (memories.lance/ with _transactions/, _versions/, data/) but never creates the actual .lance table files
2. No startup check logs: Despite `setTimeout(() => void runStartupChecks(), 0)` being called in the register() function, no "initialized successfully" or "startup checks failed" messages appear in OpenClaw logs
3. Database wiped on restart: When gateway restarts, the entire memories.lance directory is cleared
4. Status always "unavailable": The OpenClaw doctor.memory.status probe consistently returns unavailable

Manual Workaround Attempted

I manually created the LanceDB tables using the plugin's exact schema:

const schemaEntry = {
  id: '__schema__',
  text: '',
  vector: Array(768).fill(0),  // nomic-embed-text = 768 dims
  category: 'other',
  scope: 'global',
  importance: 0,
  timestamp: 0,
  metadata: '{}'
};

The table was created successfully with .lance data files present. But on gateway restart, the database was wiped again.

Diagnosis

The symptoms suggest one of these issues:

1. Silent initialization failure: The MemoryStore.doInitialize() may be failing silently — possibly due to the delete('id = "schema"') call (which throws "not yet implemented" in this LanceDB version). But this error should be caught and logged, and it isn't.

2. Fire-and-forget timing issue: The `runStartupChecks()` is scheduled via `setTimeout(..., 0)` (fire-and-forget). If the plugin's internal health check runs before the startup checks complete, it sees an uninitialized store and marks the plugin unavailable. The startup check logging may also be going to a different log sink than OpenClaw's main logs.

3. Database wipe on restart: There may be code in the plugin or OpenClaw core that clears the LanceDB database directory on startup under certain conditions.

Plugin Config Used

{
  "embedding": {
    "apiKey": "${OPENAI_API_KEY}",
    "model": "nomic-embed-text",
    "baseURL": "http://localhost:11434/v1",
    "dimensions": 768,
    "chunking": true
  },
  "autoCapture": true,
  "autoRecall": false,
  "dbPath": "/Users/luna/.openclaw/memory/lancedb-pro",
  "retrieval": {
    "mode": "hybrid",
    "rerank": "none",
    "vectorWeight": 0.7,
    "bm25Weight": 0.3,
    "minScore": 0.3
  }
}

Questions

• Is there a `delete` operation on the database directory in the plugin or OpenClaw core on startup?
• Should the `runStartupChecks()` logging go through a different sink than `api.logger`?
• Is there a known incompatibility with the parameterized delete syntax `delete('id = ?', [value])` in certain LanceDB versions?

Request

• Fix the database wipe on restart
• Ensure startup check errors are properly logged
• Add more verbose logging to doInitialize() to surface silent failures

[AliceLJY]

Thanks for the thorough writeup, @DSchexy. A few things to check based on the symptoms you're describing:

Re: database wiped on restart

This is the most concerning part. The plugin itself does NOT have code that deletes the database directory on startup. However, there are two known scenarios that can cause this:

1. LanceDB version mismatch with delete() syntax — you mentioned @lancedb/lancedb 1.x. The doInitialize() path runs a delete('id = "__schema__"') call after writing the schema probe entry. In some LanceDB 1.x builds on macOS arm64, a failed parameterized delete can corrupt the transaction log, and LanceDB's next open interprets the corrupted log as "empty database". Can you check: after a restart, does memories.lance/_transactions/ still have files, or is it truly empty?

2. OpenClaw gateway --clean or --reset-plugins flag — some startup scripts pass these flags which trigger a full plugin data wipe. Check your gateway startup command/script.

Re: startup checks never log

The setTimeout(() => void runStartupChecks(), 0) is intentionally fire-and-forget, but the logs should still appear via api.logger. Two things to verify:

• Check if your OpenClaw log level is set to info or lower (the startup checks log at info level)
• If using a process manager (pm2, systemd), the early logs might be captured before the log sink is fully attached

Re: status always "unavailable"

This follows from the above — if the store never successfully initializes (or gets wiped), the health probe returns unavailable.

Suggested debug steps:

# 1. Check if delete syntax is the culprit — run with DEBUG logging
OPENCLAW_LOG_LEVEL=debug openclaw start

# 2. Check transaction log state after restart
ls -la ~/.openclaw/memory/lancedb-pro/memories.lance/_transactions/

# 3. Check if gateway is passing clean flags
ps aux | grep openclaw

I'll also add more verbose logging to doInitialize() in the next patch release to make these failures visible. Let me know what you find from the above steps.

[DSchexy]

Thanks for the debugging steps! Here's what I found:

Database persistence — NOT the issue:
Database survives full gateway restart with 2 rows intact (schema row + 1 captured memory entry). Transaction logs survive. No --clean flags on gateway process.

Ollama embedder — works perfectly:
curl to /api/embeddings returns 768-dim vector in <1s.

Health probe still "unavailable" even though:

• LanceDB database has data
• Ollama embedder works
• Plugin captures memories

Questions:

1. What specifically does the health probe check? Store + retrieve test memory?
2. Race condition where health probe runs before ensureInitialized() completes?
3. Is doInitialize() actually async? lancedb.connect() isn't in a top-level try/catch.
4. Could the health probe be failing due to a timeout or specific check that isn't logged?

Database and embedder work fine in isolation. "Unavailable" seems to come from health probe specifically.