[Priority 3] Add access control and governance layer

[marknutter]

Problem

No way to mark memories as:

• Private (don't include in shared contexts)
• Sensitive (require human approval before loading)
• Excluded from recall (manual-access-only)
• Redacted fields (API keys, passwords)

All memory entries are treated equally regardless of content sensitivity.

Proposal (from AIGNE paper analysis)

Extend memory schema with access metadata:

{
  "id": "m_sensitive_data",
  "summary": "Production API keys",
  "tags": ["credentials", "prod"],
  "access": {
    "visibility": "private",
    "exclude_from": ["recall"],
    "require_human_approval": true,
    "redact_fields": ["api_key", "password"]
  }
}

Enforce in:

• Search: skip entries with exclude_from: ["recall"]
• Display: redact sensitive fields unless approved
• Recall: prompt for confirmation before loading require_human_approval entries

Implementation

1. Add access JSON field to memory schema
2. Update search logic to respect exclude_from
3. Add redaction logic to memory_extract
4. CLI: rlm memory set-access <entry_id> --visibility private
5. User prompts for approval when needed

Impact

• Security (prevent accidental leakage)
• Privacy (GDPR/HIPAA compliance)
• Trust (explicit control over sensitive data)
• Multi-user support (team vs private memories)

Effort

2-3 days

Related

• Access control/governance from 'Everything is Context' paper
• Metadata-driven permissions