Memory Tools — Enhanced Search for OpenClaw

Problem

OpenClaw's memory-lancedb plugin uses OpenAI embeddings only and has weak full-text search for non-English languages (no stemming, no morphology). For Russian and other inflected languages, recall is poor.

Solution

Python scripts that build a parallel vector+FTS5 index over your chat history:

1. build_vector_index.py — Chunks all LCM chat history, embeds via Google gemini-embedding-2-preview (fallback: Ollama nomic-embed-text), stores in SQLite
2. search_history_fast.py — Hybrid search combining vector similarity + FTS5 lexical scoring with numpy memmap for speed
3. unified_search.py — Single entry point that searches vector DB + entity graph + LCM memory
4. search_history.sh — Shell wrapper for quick CLI use

Requirements

• Python 3.10+
• numpy
• Google AI API key (for embeddings) OR local Ollama with nomic-embed-text
• OpenClaw with LCM database (~/.openclaw/lcm.db)

Installation

# Place in your OpenClaw workspace
cp -r memory-tools/ ~/.openclaw/workspace/scripts/memory/

# Install dependencies
pip3 install numpy

# Build the index (first run takes a while)
python3 build_vector_index.py

# Search
./search_history.sh "your query here" 10

Architecture

lcm.db (chat history)
  │
  ├─ build_vector_index.py ──→ chat_vectors.db (SQLite + FTS5)
  │     Google embeddings (primary)
  │     Ollama embeddings (fallback)
  │
  ├─ search_history_fast.py ──→ Hybrid vector+lexical search
  │     numpy memmap for fast cosine similarity
  │     FTS5 for morphology-aware text matching
  │     Score fusion: 0.65 * vector + 0.35 * lexical
  │
  └─ unified_search.py ──→ Multi-source orchestrator
        vector DB + entity graph + raw LCM

Configuration

Set via environment variables:

Variable	Default	Description
OLLAMA_HOST	http://localhost:11434	Ollama server URL
GOOGLE_AI_API_KEY	from openclaw.json	Google AI API key for embeddings

Key Numbers

• ~100K chunks indexed from chat history
• 3072-dimensional Google embeddings (1024 for Ollama fallback)
• Hybrid scoring: 65% vector + 35% lexical
• Search latency: <2s typical (with numpy memmap)