Feature Request: PII Redaction Layer

Problem Statement

OpenClaw's power comes from deep integration with users' personal data — emails, calendars, files, contacts, financial information. All of this context is sent unredacted to LLM providers. Users who follow the recommended patterns (USER.md with personal details, MEMORY.md with life context) are sending sensitive information to third-party APIs with no automatic protection.

This is a trust and adoption blocker, especially for privacy-conscious users and enterprise deployments.

Proposed Solution

Pre-Processing Redaction Layer

• Sits between the agent and the LLM provider in the gateway.
• Scans all outgoing context for PII patterns.
• Replaces detected PII with reversible placeholders: [PHONE_1], [EMAIL_2], [ADDRESS_1].

Supported Categories

• Phone numbers (international formats)
• Email addresses
• Physical addresses
• Social Security Numbers / national IDs
• Credit card numbers
• Bank account numbers
• Dates of birth
• IP addresses
• API keys and tokens

Reversible Mapping

• Local-only mapping file stores placeholder → original pairs.
• Restoration happens at tool execution time: when the agent needs to send an actual email or make a call, the real value is injected.
• Mapping file is session-scoped and never sent to providers.

Configuration

pii:
  enabled: true
  categories:
    phone: true
    email: true
    address: true
    financial: true    # SSN, credit cards, bank accounts
    api_keys: true
  exclude_providers:
    - local/ollama     # no need to redact for local models

Prior Art

A community skill (pii-redaction) already exists with regex-based detection and redact/restore modes. Promoting this to a core gateway feature would increase adoption and ensure consistent protection.

User Impact

• Privacy: Users can share personal context freely, knowing sensitive data is scrubbed before reaching providers.
• Trust: Lowers the barrier for users hesitant to put real personal data in USER.md and MEMORY.md.
• Compliance: Helps with GDPR, CCPA, and other data protection requirements.
• Enterprise: A prerequisite for organizational deployments where data handling policies are strict.

Technical Considerations

• Detection accuracy: Regex-based detection is fast but has false positives/negatives. Hybrid approach (regex + lightweight NER) improves accuracy.
• Performance: Must add <50ms latency per message. Regex scanning is typically <5ms.
• Restoration timing: Critical to restore PII at the right moment — before tool execution (sending email), not before LLM processing.
• Context coherence: The LLM must still be able to reason about redacted data structurally ("call [PHONE_1]" should work).
• Existing skill: The pii-redaction skill in the community provides a solid starting point for the regex patterns and restore logic.

Priority

MEDIUM. Important for trust, adoption, and compliance. Not blocking current users but would significantly expand the addressable user base. Low implementation cost given the existing community skill.