LUCID

Autonomous memory retrieval for OpenClaw agents

Hooks into OpenClaw's native ContextEngine API — your agent searches its knowledge base before the model ever sees your message.

git clone https://github.com/Spaztazim/lucid-context-engine.git ~/.openclaw/extensions/lucid-context-engine
cd ~/.openclaw/extensions/lucid-context-engine && npm install && npm run build

Standalone scoring libraries compute relevance — but you still have to wire up every part of the retrieval pipeline yourself on every message. That's plumbing no one should write twice.

LUCID handles the full pipeline. Install once. It runs automatically on every turn.

Every time you send a message, LUCID:

1. Filters trivially — skips search on "ok", "thanks", heartbeats, and short acks
2. Searches your workspace — hybrid BM25 + semantic search via QMD
3. Scores by salience — not just relevance, but recency, file type, and collection priority
4. Respects your budget — injects only what fits in your remaining context window
5. Passes it through — the model sees recalled context as part of its system prompt, automatically

No prompt changes. No tool calls. No manual retrieval. It just works.

flowchart LR
    A([User Message]) --> B{Trivial?}
    B -- yes --> G([Pass Through])
    B -- no --> C[QMD Hybrid Search\nBM25 + Semantic]
    C --> D[Salience Scoring\nrelevance × recency\n× type × collection]
    D --> E{Budget\nCheck}
    E --> F[Context Injection\nSystem Prompt Addition]
    F --> H([Model])

salience = qmd_score × recency_weight × type_weight × collection_weight

git clone https://github.com/Spaztazim/lucid-context-engine.git ~/.openclaw/extensions/lucid-context-engine

Then activate in your OpenClaw config (~/.openclaw/config.json or per-agent):

{
  "plugins": {
    "slots": {
      "contextEngine": "lucid"
    },
    "entries": {
      "lucid-context-engine": {
        "enabled": true
      }
    }
  }
}

Restart your agent. Done.

All options are optional — defaults work out of the box.

{
  "plugins": {
    "entries": {
      "lucid-context-engine": {
        "enabled": true,

        // Max results to inject per turn
        "topK": 5,

        // Minimum salience score (0.0–1.0) to include a result
        "threshold": 0.3,

        // Path to qmd-shim.js (auto-detected if omitted)
        "qmdShimPath": "~/clawd/tools/qmd-shim.js",

        // Max milliseconds to wait for QMD before falling back
        "timeoutMs": 5000
      }
    }
  }
}

• OpenClaw v3.7+
• QMD workspace search daemon (ships with OpenClaw)
• Node.js 18+

npm install
npm run build

Output goes to dist/.