lobs-memory

Persistent memory search server with neural reranking and hybrid search (BM25 + vector).

Features

• Hybrid search: BM25 keyword search + vector semantic search
• Neural reranking: Cross-encoder reranking for higher quality results
• Persistent models: Keeps embedding and reranker models loaded in memory
• File watching: Automatically re-indexes changed files
• Embedding cache: Avoids re-embedding unchanged text
• MMR diversity: Ensures result variety
• Temporal decay: Prioritizes recent files (configurable)

Requirements

• Bun runtime
• LM Studio running locally on localhost:1234 with text-embedding-qwen3-embedding-4b loaded
• Reranker GGUF (optional): ~/.cache/qmd/models/hf_ggml-org_qwen3-reranker-0.6b-q8_0.gguf

Installation

bun install

Configuration

Edit config.json or set environment variables:

{
    "port": 7420,
    "lmstudio": {
      "baseUrl": "http://localhost:1234/v1",
      "embeddingModel": "text-embedding-qwen3-embedding-4b"
    },
  "models": {
    "reranker": "~/.cache/qmd/models/hf_ggml-org_qwen3-reranker-0.6b-q8_0.gguf"
  },
  "collections": [
    {
      "name": "memory",
      "path": "~/.lobs/workspace",
      "pattern": ["MEMORY.md", "memory/**/*.md"]
    }
  ]
}

Environment variables:

• PORT - Server port (default: 7420)
• LMSTUDIO_URL - LM Studio base URL
• EMBEDDING_MODEL - Embedding model name
• RERANKER_MODEL - Path to reranker GGUF

Usage

Start server

bun run start
# or with auto-reload during development:
bun run dev

Search API

curl -X POST http://localhost:7420/search \
  -H "Content-Type: application/json" \
  -d '{"query": "github issues", "maxResults": 5}'

Response:

{
  "results": [
    {
      "path": "/path/to/file.md",
      "startLine": 10,
      "endLine": 50,
      "score": 0.85,
      "snippet": "...",
      "source": "memory",
      "citation": "/path/to/file.md:10-50"
    }
  ],
  "query": "github issues",
  "timings": {
    "totalMs": 150,
    "bm25Ms": 2,
    "vectorMs": 45,
    "rerankMs": 98
  }
}

Health check

curl http://localhost:7420/health

Status

curl http://localhost:7420/status

Manual re-index

curl -X POST http://localhost:7420/index

Architecture

• server/config.ts - Configuration loading
• server/db.ts - SQLite database with FTS5 and sqlite-vec
• server/chunker.ts - Markdown chunking
• server/embedder.ts - LM Studio embedding client
• server/reranker.ts - node-llama-cpp reranker
• server/search.ts - Full search pipeline
• server/indexer.ts - File indexing and watching
• server/index.ts - HTTP server

Search Pipeline

1. BM25 search - Keyword search via SQLite FTS5
2. Vector search - Semantic search via sqlite-vec
3. Weighted merge - Combine scores (default: 70% vector, 30% text)
4. Neural reranking - Cross-encoder scoring (if available)
5. Temporal decay - Exponential decay based on file date
6. MMR diversity - Maximal marginal relevance filtering
7. Return top-K - Default 8 results

Performance

• Cold start: ~3-5s (loads reranker GGUF)
• Search latency: 100-200ms (with reranking)
• Embedding: ~25ms per chunk (LM Studio)
• Reranking: ~50-100ms for 20 candidates

Development

# Install dependencies
bun install

# Run with auto-reload
bun run dev

# Check database
sqlite3 ~/.lobs/plugins/lobs-memory/index.db "SELECT COUNT(*) FROM documents;"

Troubleshotics

LM Studio connection errors:

• Make sure LM Studio is running on localhost:1234
• Check that text-embedding-qwen3-embedding-4b is loaded in LM Studio

Reranker not loading:

• Check that the GGUF file exists at the configured path
• The server will work without reranking if the model is missing (degrades gracefully)

Slow initial startup:

• First run indexes all files (can take a few minutes for large collections)
• Subsequent starts are fast (~3s) because files are cached
• The server is responsive immediately; indexing runs in background

License

Private - Part of the lobs-ai project.