Summary

chunks_vec not updated — sqlite-vec unavailable. Vector recall degraded. is logged even when the sqlite-vec extension successfully loads. The warning fires whenever vectorReady === false, which can be caused by reasons unrelated to sqlite-vec (e.g., missing embedding provider). This makes diagnosing the real cause very hard.

Environment

• OpenClaw 2026.4.27
• Node.js v25.9.0 (built-in node:sqlite, SQLite 3.51.3)
• macOS arm64
• sqlite-vec 0.1.9 (with sqlite-vec-darwin-arm64 binary present)

Reproduction

1. Configure OpenClaw with claude-cli runtime, no OpenAI API key for embeddings (only ChatGPT / openai-codex OAuth, which does not grant embedding API access).
2. Run openclaw memory index --agent <id> --verbose --force
3. Observe: chunks_vec not updated — sqlite-vec unavailable. Vector recall degraded. Further duplicate warnings suppressed.

Investigation

Patching dist/engine-storage-CO3qFXo7.js loadSqliteVecExtension with console.error at each step prints:

[DBG] vec start, extPath= undefined
[DBG] vec module loaded
[DBG] vec extPath= .../sqlite-vec-darwin-arm64/vec0.dylib
[DBG] enableLoadExtension OK
[DBG] vec.load OK

→ sqlite-vec extension is successfully loaded. The "unavailable" warning still fires immediately after.

Verified independently with a direct Node.js test:

const sqlite = await import('node:sqlite');
const db = new sqlite.DatabaseSync(':memory:', { allowExtension: true });
db.enableLoadExtension(true);
const sv = await import('sqlite-vec');
sv.load(db);
db.prepare('SELECT vec_version() as v').get();
// → { v: 'v0.1.9' }  ✅

Root cause

manager-jzSMQjEt.js:1843:

function logMemoryVectorDegradedWrite(params) {
  if (!params.vectorEnabled || params.vectorReady || params.chunkCount <= 0 || params.warningShown) return params.warningShown;
  const errDetail = params.loadError ? `: ${params.loadError}` : "";
  params.warn(`chunks_vec not updated — sqlite-vec unavailable${errDetail}. Vector recall degraded. Further duplicate warnings suppressed.`);
  return true;
}

The condition fires whenever vectorReady === false. vectorReady requires a valid dimensions param from a successful embedding call (ensureVectorReady, line 1031). Without a working embedding provider, dimensions is undefined, the vector table is never created, and the warning falsely blames sqlite-vec — even though loadSqliteVecExtension succeeded and loadError is null.

This is likely the same surface symptom reported in #65156, but the underlying mechanism is different (and that issue's ABI-mismatch hypothesis does not apply here).

Suggested fix

Distinguish the warning message:

• Only mention sqlite-vec when loadVectorExtension() itself failed (i.e., vector.loadError is set).
• When the extension loaded but vectorReady is false because no embedding dimensions were ever resolved, surface that instead — e.g., chunks_vec not updated — embedding provider unavailable / dimensions undetermined. Vector recall degraded.

Workaround

Configure a valid embedding provider (OpenAI API key, Voyage, Mistral, etc.) under agents.defaults.memorySearch. Note that openai-codex / ChatGPT subscription OAuth does not provide embedding API access.