Bug Summary

memory_search tool returns empty results despite:

• Database fully populated with chunks and vectors
• OpenAI embeddings working correctly
• FTS and vector search both functional when tested manually
• No errors logged

Environment

• OpenClaw version: 2026.2.3-1
• OS: Linux 6.17.0-12-generic (x64)
• Memory backend: openai (text-embedding-3-small)
• Vector storage: sqlite-vec (v0.1.7-alpha.2)

Steps to Reproduce

1. Index memory files with openclaw memory index --force
2. Verify status shows "Indexed: 5/5 files · 21 chunks"
3. Run memory_search with any query:

   {"query": "Rock RMS mobile app"}

4. Result: Empty array [] with no errors

Expected Behavior

Search should return relevant chunks from indexed memory files.

Actual Behavior

{
  "results": [],
  "provider": "openai",
  "model": "text-embedding-3-small",
  "citations": "auto"
}

Silent failure - no errors in logs, tool reports success but returns no matches.

Diagnostic Data

Database Status (All Healthy)

SELECT COUNT(*) FROM files;        -- 5 files
SELECT COUNT(*) FROM chunks;       -- 21 chunks  
SELECT COUNT(*) FROM chunks_vec;   -- 21 vectors (via vec0 extension)
SELECT COUNT(*) FROM chunks_fts;   -- 21 rows

Manual Tests (All Pass)

1. OpenAI API Test:

curl https://api.openai.com/v1/embeddings \
  -H "Authorization: Bearer sk-..." \
  -d '{"input":"Rock RMS","model":"text-embedding-3-small"}'

✅ Returns valid 1536-dim embedding

2. FTS Search Test:

SELECT * FROM chunks_fts WHERE chunks_fts MATCH 'Rock';

✅ Returns 11 matching rows with snippets

3. Vector Similarity Test (Python):

# Using sqlite-vec extension with vec_distance_cosine()
# Query: "Rock RMS mobile"
# Returns:
#   Distance: 0.4544 | Path: memory/rock_rms_mobile_guide.md
#   Distance: 0.5282 | Path: memory/rock_rms_mobile_guide.md
#   ...

✅ Returns 5 relevant chunks with cosine distances

Configuration

{
  "agents": {
    "defaults": {
      "memorySearch": {
        "enabled": true,
        "sources": ["memory", "sessions"],
        "provider": "openai"
      }
    }
  }
}

Vector Table Schema

CREATE VIRTUAL TABLE chunks_vec USING vec0(
  id TEXT PRIMARY KEY,
  embedding FLOAT[1536]
);

All 21 chunks have valid embeddings stored as JSON arrays.

Root Cause Analysis

The memory_search tool appears to be failing to execute the actual search query, despite:

• ✅ Correctly identifying provider (openai)
• ✅ Loading the database
• ✅ Having valid authentication
• ✅ Having indexed data

The tool returns empty results instead of executing either:

1. FTS query against chunks_fts table
2. Vector similarity query against chunks_vec table

Both work perfectly when tested manually.

Related Issues

Possibly related to:

• QMD memory_search silently returns empty in Discord guild channels (default scope denies chatType 'channel') #10191 (QMD scope issues causing empty results)
• [Bug]: Unable to use QMD without OpenAI auth being configured #9143 (Memory search fails silently after auth error)
• memory_search with qmd backend: collection name mismatch causes empty results #9854 (Collection name mismatch)

Workaround

Direct file reading works. Database is accessible and functional — just not through the memory_search tool.

Suggested Fix

Add debug logging to the memory search tool to expose where the query execution fails. Current silent failure makes debugging impossible.