openclaw - 💡(How to fix) Fix v2026.3.28: memory-lancedb fails to load (missing dist/package.json + @lancedb/lancedb) and recall broken with embedding context length error [1 comments, 2 participants]

Environment

• OpenClaw version: 2026.3.28 (upgraded from 2026.3.24)
• Platform: macOS 26.4 arm64 (Apple Silicon Mac mini)
• Embedding model: mxbai-embed-large via Ollama

Issues

1. memory-lancedb fails to load on startup

After upgrading to 2026.3.28, memory-lancedb fails to load with:

Error: ENOENT: no such file or directory, open '/opt/homebrew/lib/node_modules/openclaw/dist/package.json'

Root cause: lancedb-runtime-*.js reads ./package.json (relative to its location in dist/) to find the @lancedb/lancedb dependency spec. But dist/package.json does not exist in the 2026.3.28 npm install — only the root package.json does.

Workaround that fixed it:

1. Copied root package.json to dist/package.json
2. Added "@lancedb/lancedb": "^0.27.1" to its dependencies
3. Also ran npm install in dist/extensions/memory-lancedb/ to install the dep there

Then gateway restart loaded memory-lancedb successfully.

2. memory-lancedb recall fails with context length error

After the load issue was fixed, recall fails on every inbound message:

memory-lancedb: recall failed: Error: 400 the input length exceeds the context length

The recall query (built from the inbound message including Discord metadata) exceeds the embedding model's context window (512 tokens for mxbai-embed-large). The captureMaxChars: 400 config option truncates stored text but does not appear to truncate the recall query.

This worked correctly on 2026.3.24. Something in the recall path changed in 2026.3.28.

Config:

{
  "memory-lancedb": {
    "enabled": true,
    "config": {
      "embedding": {
        "apiKey": "ollama",
        "baseUrl": "http://localhost:11434/v1",
        "model": "mxbai-embed-large",
        "dimensions": 1024
      },
      "autoCapture": false,
      "autoRecall": true,
      "captureMaxChars": 400
    }
  }
}

Impact

Both issues together mean memory recall is non-functional after upgrading to 2026.3.28. The startup ENOENT blocks loading entirely, and even after the workaround, recall fails on every turn.