Semantic Memory 🧠

Search your memories by meaning, not keywords. Uses vector embeddings to find relevant information even when you don't remember the exact words.

Built on ChromaDB for fast, private, local vector search.

Setup

# Index existing memories
python3 ~/.openclaw/semantic-memory/semantic_memory.py index

Usage

# Index all memory files (run after installing or to refresh)
python3 ~/.openclaw/semantic-memory/semantic_memory.py index

# Search memories by meaning
python3 ~/.openclaw/semantic-memory/semantic_memory.py search "what did we decide about the trading bot"

# Add a new memory
python3 ~/.openclaw/semantic-memory/semantic_memory.py add "Remember to check the OANDA bot logs daily"

# Show stats
python3 ~/.openclaw/semantic-memory/semantic_memory.py stats

How It Works

1. Indexing — Reads all .md files from ~/.openclaw/workspace/memory/, generates vector embeddings via Gemini API, stores in ChromaDB

1. Search — Converts your query to a vector, finds most similar memories using cosine similarity

1. Results — Returns relevant memories ranked by semantic similarity

Examples

Before (keyword search)

Query: "GBP USD trades"

Results: Only exact matches for "GBP USD"

After (semantic search)

Query: "What pairs did we trade on OANDA?"

Results: Finds GBP/USD, EUR/USD, USD/JPY etc. even without exact phrase match

Requirements

• ChromaDB — Local vector database (pip install chromadb)
• Gemini API key — For generating embeddings (optional, falls back to text search)
• Get key at: https://makersuite.google.com/app/apikey
• Save to: ~/.openclaw/credentials/gemini.json as {"api_key": "YOUR_KEY"}

Without Gemini key, uses simple text search as fallback.

Memory Sources

Automatically indexes:

• ~/.openclaw/workspace/memory/*.md — Daily memory files
• Manual adds via add command

Files

~/.openclaw/semantic-memory/
├── semantic_memory.py   # Main script
├── memory.log           # Log file
└── data/                # ChromaDB storage

Integration

Add to cron for automatic indexing:

# Re-index daily at 4am
0 4 * * * python3 ~/.openclaw/semantic-memory/semantic_memory.py index

Or call from other skills to search memories:

import subprocess
result = subprocess.run(
    ['python3', '/home/umbrel/.openclaw/semantic-memory/semantic_memory.py', 
     'search', 'trading decisions'],
    capture_output=True, text=True
)

Why This Matters

Regular search: "找 exactly this word"

Semantic search: "找 this meaning"

Even if I don't remember "OANDA bot flip setting", I might find "bot was losing because FLIP was disabled" — semantic search bridges that gap.

Dependencies

• chromadb — Vector database (installed with pip)
• gemini API key — For embeddings (optional)
• Python 3.8+