BEE Belief Extraction Engine

---

name: openclaw-bee description: "Install and configure BEE — the Belief Extraction Engine for OpenClaw. Gives agents persistent structured memory across sessions. Auto-extracts beliefs at session end, scopes by agent namespace, and injects recalled context on agent start. Use when: setting up agent memory, configuring belief persistence, or troubleshooting BEE. NOT for: general memory questions (use built-in memory tools instead)." homepage: https://github.com/skysphere-labs/openclaw-bee metadata: { "openclaw": { "emoji": "🐝", "requires": { "bins": ["npm"] } } }

---

BEE — Belief Extraction Engine

Give your OpenClaw agents persistent, structured memory across sessions.

What BEE Does

BEE hooks into the OpenClaw lifecycle and:

Extracts beliefs at session end via a lightweight LLM call (Haiku by default)

Injects recalled context before every agent spawn

Scopes by namespace — each agent (VECTOR, FORGE, ORACLE, etc.) has isolated beliefs

Deduplicates — cosine similarity check prevents duplicate beliefs (>0.92 → merge)

Tracks spawns — monitors subagent budget per session

Beliefs live in a SQLite database (vector.db) and persist indefinitely across restarts.

---

Installation

Step 1 — Install the package

From npm (recommended):

npm install -g @skysphere-labs/openclaw-bee

From GitHub (latest):

npm install -g github:skysphere-labs/openclaw-bee

Step 2 — Configure openclaw.json

Add BEE to your extensions in ~/.openclaw/openclaw.json:

{
  "extensions": {
    "entries": {
      "bee": {
        "enabled": true,
        "config": {
          "dbPath": "~/.openclaw/workspace/state/vector.db",
          "agentId": "main",
          "extractionEnabled": true,
          "extractionModel": "anthropic/claude-haiku-4-5",
          "maxCoreBeliefs": 10,
          "maxActiveBeliefs": 5,
          "maxRecalledBeliefs": 5
        }
      }
    }
  }
}

Step 3 — Restart the gateway

openclaw gateway restart

BEE will run its database migration on first start and begin capturing beliefs.

---

Configuration Options

| Option | Default | Description | |--------|---------|-------------| | dbPath | *required* | Path to your SQLite database | | agentId | "main" | Namespace for belief scoping | | extractionEnabled | true | Enable/disable belief extraction | | extractionModel | "anthropic/claude-haiku-4-5" | Model used for extraction (cheapest works well) | | extractionMinConfidence | 0.55 | Minimum confidence to store a belief (0-1) | | maxCoreBeliefs | 10 | Core beliefs injected into every session | | maxActiveBeliefs | 5 | Recently active beliefs injected | | maxRecalledBeliefs | 5 | Semantically recalled beliefs per query | | maxOutputChars | 2000 | Max chars of belief context injected | | debug | false | Enable verbose logging | | spawnBudgetWarning | 20 | Warn when subagent spawns exceed this threshold |

---

Verifying It Works

After restart, ask your agent: > "How many beliefs do you have?"

Or check directly:

sqlite3 ~/.openclaw/workspace/state/vector.db "SELECT COUNT(*) FROM beliefs"

You should see beliefs accumulate after sessions complete.

---

Multi-Agent Setup

For setups with multiple named agents (VECTOR, FORGE, ORACLE, etc.), use different agentId values per agent spawn. BEE scopes beliefs by agentId so each PM has isolated memory.

---

Source

GitHub: https://github.com/skysphere-labs/openclaw-bee

npm: https://www.npmjs.com/package/@skysphere-labs/openclaw-bee

Built by Skysphere AI Labs