LUCID Context Engine
Autonomous context engine plugin for OpenClaw - pre-search memory retrieval with salience scoring, QMD integration, and cross-agent memory support
Install
openclaw plugins install clawhub:lucid-context-engine
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.
? Quick Install
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
Restart your gateway after install. That's it.
The Problem
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.
How LUCID Works
Every time you send a message, LUCID:
- Filters trivially � skips search on "ok", "thanks", heartbeats, and short acks
- Searches your workspace � hybrid BM25 + semantic search via QMD
- Scores by salience � not just relevance, but recency, file type, and collection priority
- Respects your budget � injects only what fits in your remaining context window
- 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.
Architecture
Salience Formula
salience = qmd_score � recency_weight � type_weight � collection_weight
| Factor | Values |
|---|---|
| Recency | =7d: 1.5� � =30d: 1.2� � =90d: 1.0� � older: 0.8� |
| Type | LESSONS.md: 2.0� � decision: 1.5� � memory/*: 1.0� � log: 0.7� |
| Collection | memory: 1.5� � codex: 1.2� � default: 1.0� |
Features
| ?? Hybrid Search | BM25 + semantic via QMD � finds exact matches and conceptual matches |
| ?? Salience Scoring | Recency, file type, and collection priority all factor in |
| ?? Budget Aware | Respects your context window � injects top-K up to remaining tokens |
| ? Trivial Filtering | Short acks and heartbeats skip search entirely |
| ?? Cross-Collection | Reads from any QMD-indexed collection in your workspace |
| ??? Graceful Fallback | QMD unavailable? Silent passthrough � no errors, no interruptions |
| ?? Zero Config | Auto-detects QMD path, sensible defaults out of the box |
Install
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.
Configuration
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
}
}
}
}
| Option | Default | Description |
|---|---|---|
topK | 5 | Max results injected per turn |
threshold | 0.3 | Minimum salience score to include |
qmdShimPath | auto | Path to qmd-shim.js |
timeoutMs | 5000 | QMD timeout in milliseconds |
Comparison
| LUCID | Standalone Scoring Library | |
|---|---|---|
| Full retrieval pipeline | ? | ? Manual wiring required |
| Automatic on every turn | ? | ? Must call per-message |
| Context budget management | ? | ? |
| Trivial prompt filtering | ? | ? |
| Graceful fallback | ? | ? |
| OpenClaw native integration | ? | ? |
| Lines of setup code | ~5 | ~100+ |
Requirements
- OpenClaw v3.7+
- QMD workspace search daemon (ships with OpenClaw)
- Node.js 18+
Building from Source
npm install
npm run build
Output goes to dist/.
Built By
Almost Spec Labs
github.com/Spaztazim