vecmem

Semantic memory CLI with vector embeddings for Japanese. Small enough to run on a Raspberry Pi.

Install

uv pip install git+https://github.com/nagayama/vecmem.git

Usage

# Add a memory
vecmem add --summary "TDDで開発するのが好き" --details "Red-Green-Refactorサイクル"

# Search by meaning
vecmem search --query "テスト駆動開発"

# List recent memories
vecmem list --limit 10

# Update
vecmem update --id 1 --summary "新しい要約"

# Delete
vecmem delete --id 1

Daemon Mode

Keep the embedding model in memory for faster responses (cold start 5-15s → warm ~ms).

# Start daemon in background
vecmem daemon start

# Check status
vecmem daemon status
# {"status": "running", "pid": 12345}

# Stop daemon
vecmem daemon stop

Use DaemonClient from vecmem.daemon.client to communicate with the running daemon.

Output

All commands output JSON:

// add
{"status": "ok", "id": 1}

// search
{"results": [{"id": 1, "summary": "...", "distance": 0.23}]}

// list
{"memories": [{"id": 1, "summary": "...", "created_at": "..."}]}

Data

• Database: ~/.vecmem/vecmem.db (VECMEM_DB_PATH)
• Socket: ~/.vecmem/vecmem.sock (VECMEM_SOCKET_PATH)
• PID file: ~/.vecmem/vecmem.pid (VECMEM_PID_PATH)

Tech

• Embedding: hotchpotch/static-embedding-japanese (128 dim)
• Vector DB: sqlite-vec