Installation

Get the Epistemic Memory Engine running in your OpenClaw instance in under 5 minutes.

Prerequisites

OpenClaw v2026.4.10 or later
Node.js 20+ (included in the Docker image)
pnpm 9+ (workspace manager)
An OpenAI API key (for embeddings — text-embedding-3-small)

Docker (Recommended)

If you're running OpenClaw via Docker Compose, the epistemic plugin is bundled automatically:

docker-compose.yml

services:
  openclaw-gateway:
    image: ghcr.io/monasprox/openclaw:latest
    volumes:
      - ./data:/app/data
    environment:
      - NODE_ENV=production

Start the stack:

docker compose up -d

From Source

# Clone the monorepo
git clone https://github.com/monasprox/epistemic.git
cd epistemic

# Install dependencies
pnpm install

# Build the plugin
pnpm --filter @openclaw/memory-claw build

Configuration

Add the plugin to your OpenClaw configuration file:

openclaw.json

{
  "plugins": {
    "entries": {
      "epistemic": {
        "enabled": true,
        "config": {
          "embedding": {
            "apiKey": "sk-your-openai-api-key",
            "model": "text-embedding-3-small",
            "dimensions": 1536
          },
          "ownerId": "your-telegram-user-id",
          "ownerAliases": ["your-zalo-id", "your-discord-id"],
          "decayEnabled": true,
          "conflictAutoResolve": false,
          "channelIsolation": true
        }
      }
    }
  }
}

Configuration Reference

Parameter	Type	Default	Description
`embedding.apiKey`	`string`	—	Required. OpenAI API key for vector embeddings.
`embedding.model`	`string`	`"text-embedding-3-small"`	Embedding model name.
`embedding.dimensions`	`number`	`1536`	Vector dimensions.
`ownerId`	`string`	—	Required. Primary owner identifier (e.g. Telegram user ID).
`ownerAliases`	`string[]`	`[]`	Alternative IDs for the same owner across channels.
`decayEnabled`	`boolean`	`true`	Enable temporal decay engine.
`conflictAutoResolve`	`boolean`	`false`	Automatically resolve conflicts (higher confidence wins).
`channelIsolation`	`boolean`	`true`	Prevent cross-channel memory leakage.
`decayClasses`	`object`	see Decay Classes	Custom half-lives per decay class.

Verify Installation

After starting OpenClaw, check the logs for:

[epistemic] ✓ LanceDB initialized at data/epistemic/lance
[epistemic] ✓ Embedding model: text-embedding-3-small (1536d)
[epistemic] ✓ 11 tools registered
[epistemic] ✓ Decay engine active (interval: 6h)

You can also call the health endpoint:

curl -X POST http://localhost:18789/api/tool \
  -H "Authorization: Bearer YOUR_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"tool": "epistemic_health", "input": {}}'

Next: Follow the Quick Start guide to store your first memory.