@biginformatics/openclaw-wagl

OpenClaw plugin that brings wagl DB-first memory to any agent.

• Auto-recall — injects relevant memories before each agent turn (before_agent_start hook)
• Auto-capture — saves the last assistant response after each turn (agent_end hook)
• Tools — wagl_recall and wagl_store available to the agent at runtime
• Memory slot — registers as memory-wagl, taking the OpenClaw memory slot

Install

openclaw plugins install @biginformatics/openclaw-wagl

Then add to openclaw.json:

{
  "plugins": {
    "slots": { "memory": "memory-wagl" },
    "entries": {
      "memory-wagl": {
        "package": "@biginformatics/openclaw-wagl"
      }
    }
  }
}

Requirements

1. wagl binary

Install wagl on the agent's host and ensure it is on PATH:

# Download the latest release binary for your platform
# https://github.com/BigInformatics/wagl/releases
cp wagl /usr/local/bin/wagl
chmod +x /usr/local/bin/wagl
wagl init   # creates ~/.wagl/memory.db

Set WAGL_DB to a stable path (recommended):

export WAGL_DB="$HOME/.wagl/memory.db"

Without WAGL_DB, wagl defaults to a path relative to $PWD which will vary.

2. Semantic recall (optional but recommended)

Semantic search requires:

• An OpenAI-compatible embeddings endpoint (e.g. Ollama on armada)
• The sqlite-vec extension (vec0.so)

# Point to your embeddings server
export WAGL_EMBEDDINGS_BASE_URL="http://your-ollama-host:11434"
export WAGL_EMBEDDINGS_MODEL="qwen3-embedding:0.6b"

# sqlite-vec: place vec0.so at a stable path and set SQLITE_VEC_PATH
# Download a pre-built vec0.so from the wagl releases page, or build from source:
#   cd wagl-cli && git submodule update --init && ./scripts/build-sqlite-vec-linux-x86_64.sh
#   cp third_party/sqlite-vec/linux-x86_64/vec0.so /usr/local/lib/
export SQLITE_VEC_PATH="/usr/local/lib/vec0.so"

Without semantic recall, the plugin still works — wagl_store persists memories and wagl_recall runs text-only matching (lower precision). The plugin degrades gracefully; it will not crash if SQLITE_VEC_PATH or embeddings are missing.

3. Systemd service (recommended for production)

If OpenClaw runs as a systemd user service, add a drop-in to persist env vars:

mkdir -p ~/.config/systemd/user/openclaw-gateway.service.d

cat > ~/.config/systemd/user/openclaw-gateway.service.d/wagl.conf << 'UNIT'
[Service]
Environment=WAGL_DB=/home/YOUR_USER/.wagl/memory.db
Environment=SQLITE_VEC_PATH=/usr/local/lib/vec0.so
Environment=WAGL_EMBEDDINGS_BASE_URL=http://your-ollama-host:11434
Environment=WAGL_EMBEDDINGS_MODEL=qwen3-embedding:0.6b
UNIT

systemctl --user daemon-reload
systemctl --user restart openclaw-gateway.service

Configuration

All config is optional — defaults work for a basic setup.

{
  "plugins": {
    "entries": {
      "memory-wagl": {
        "package": "@biginformatics/openclaw-wagl",
        "config": {
          "dbPath": "/home/user/.wagl/memory.db",
          "autoRecall": true,
          "autoCapture": true,
          "recallQuery": "who am I, current focus, working rules"
        }
      }
    }
  }
}

| Key | Default | Description | |---|---|---| | dbPath | $WAGL_DB → $WAGL_DB_PATH → ~/.wagl/memory.db | Path to wagl DB | | autoRecall | true | Inject recall context before each agent turn | | autoCapture | true | Save last assistant message after each turn | | recallQuery | "who am I, current focus, working rules" | Query used for auto-recall |

Known limitations

• vec0.so is not yet bundled with the wagl binary — see wagl#17 for the tracking issue
• Without semantic embeddings, recall uses text-only matching which has lower precision

License

Apache 2.0 — Copyright 2026 Informatics FYI, Inc.