OpenClaw Knowledge System (Sanitized Public Template)

A public, sanitized template of the raw → compiled knowledge loop we implemented internally for OpenClaw operations.

Inspired by ideas shared by @karpathy (original tweet). Thank you for pushing practical, corpus-first workflows into the mainstream.

Why this exists

Most teams lose context because notes are fragmented and synthesis is inconsistent. This template turns that into a repeatable system:

1. Capture source evidence in raw/
2. Compile reusable operational knowledge in compiled/
3. Preserve lineage links (what synthesis came from which sources)
4. Query corpus first before inventing context from scratch
5. Run periodic health checks for drift, staleness, and contradictions

Repository structure

research/knowledge-system/
├── README.md
├── index.md
├── raw/
│   └── {stream}/{YYYY-MM-DD}/
│       ├── SOURCES.md
│       └── snapshots...
└── compiled/
    └── {stream}/
        ├── overview.md
        └── runs/{YYYY-MM-DD}.md
scripts/
└── knowledge-compile.sh

What’s new in v2

• Redaction policy with practical before/after examples
• Quality gates for metadata, contradiction checks, freshness, and lineage
• Maturity model (L1→L5) to stage adoption
• Operational workflows (daily/weekly)
• Lightweight ROI metrics
• Anti-pattern catalog
• Weekly synthesis template + example artifact

Quick start

1) Create your first streams

Pick 1-3 streams (examples):

• product
• customer-support
• platform-ops

2) Ingest one day of raw evidence

Create:

• research/knowledge-system/raw/<stream>/<date>/
• SOURCES.md with metadata (source, captured_at, tags, confidence)

3) Compile it

Run:

bash scripts/knowledge-compile.sh 2026-04-04

This creates deterministic run notes with lineage anchors in compiled/*/runs/.

4) Query discipline

Before planning/execution asks, check compiled pages first, then raw evidence if needed.

5) Health checks

Recommended cadence:

• Daily: ingest + compile for active streams
• Tue/Fri: stale pages, broken links, conflicting facts, missing next actions

6) Weekly synthesis

Create one weekly artifact using templates/weekly-synthesis.md to capture:

• changed assumptions
• risks
• decisions
• unknowns
• next actions + lineage links

Agent ingestion flow (drop-in workflow)

Use this when wiring the repo into an agentic loop.

Step 1: Bootstrap once

git clone https://github.com/BradGroux/openclaw-knowledge-system-template.git
cd openclaw-knowledge-system-template

Step 2: Define streams

Create stream folders under:

• research/knowledge-system/raw/<stream>/
• research/knowledge-system/compiled/<stream>/

Start with 1–3 streams for signal quality.

Step 3: Ingest daily raw evidence

For each active stream/date:

1. Add snapshots (notes, decisions, task exports, etc.)
2. Update SOURCES.md with required metadata:
   • source
   • captured_at
   • tags
   • confidence

Step 4: Run compile pass

bash scripts/knowledge-compile.sh "$(date +%F)"

Step 5: Query protocol for agents

Before answering non-trivial requests:

1. Read compiled/<stream>/overview.md
2. Read latest compiled/<stream>/runs/<date>.md
3. Trace to raw SOURCES.md when confidence is low or conflicts appear
4. Only then synthesize response or plan

Step 6: Weekly governance

• Run Tue/Fri health checks (docs/quality-gates.md)
• Publish one weekly synthesis from templates/weekly-synthesis.md
• Track metrics in docs/metrics.md

Step 7: Public-sharing guardrail

Before publishing or syncing outside your org:

bash scripts/sanitize-scan.sh .

Then complete docs/sanitization-checklist.md and docs/public-redaction-policy.md review.

Sanitization policy (important)

This repo is designed for public sharing. Keep it safe:

• No personal emails, phone numbers, IDs, or tokens
• No private org names unless already public
• No local machine paths (/Users/..., /home/...)
• No internal hostnames, ports, or private URLs
• No customer data, credentials, secrets, or proprietary prompts

See docs/sanitization-checklist.md for a pre-publish checklist.

Included artifacts

• research/knowledge-system/README.md – loop and cadence
• research/knowledge-system/index.md – stream index and health checklist
• scripts/knowledge-compile.sh – sample ingestion/compile script (sanitized)
• scripts/sanitize-scan.sh – pre-publish quick sensitive-pattern scanner
• docs/public-redaction-policy.md – what to strip/replace before publishing
• docs/quality-gates.md – metadata, contradiction, freshness, lineage gates
• docs/maturity-model.md – L1→L5 adoption path
• docs/workflows.md – practical daily/weekly command flows
• docs/metrics.md – measurable ROI indicators
• docs/anti-patterns.md – common failure modes
• templates/weekly-synthesis.md – reusable weekly digest template
• examples/ – sample raw, compiled, and weekly synthesis files

License

MIT (see LICENSE).

Credits

• Conceptual inspiration: Andrej Karpathy
• OpenClaw ecosystem and community: openclaw
• Peter Steinberger: @steipete
• Operational adaptation and template packaging: Digital Meld / OpenClaw workflow patterns