Skip to content

Compliance

Varun Pratap Bhardwaj edited this page Mar 17, 2026 · 2 revisions

Compliance

SuperLocalMemory is designed for regulatory compliance from the ground up. This page covers EU AI Act, GDPR, data retention, access control, and audit capabilities.

EU AI Act (Regulation 2024/1689)

The EU AI Act takes full effect August 2, 2026. To the best of our knowledge, no existing agent memory system addresses EU AI Act compliance.

Mode A & B: Full Compliance by Architecture

Mode A operates as a zero-LLM retrieval system. Mode B adds a local LLM via Ollama. In both modes, all memory operations — storage, encoding, retrieval, and lifecycle management — execute locally without any cloud dependency.

Requirement Mode A Mode B Mode C
Data sovereignty (Art. 10) Pass Pass Requires DPA
Right to erasure (GDPR Art. 17) Pass Pass Pass
Transparency (Art. 13) Pass Pass Pass
No network calls during memory ops Yes Yes No

Key compliance points for Mode A/B:

  • Data sovereignty: No personal data leaves the device during any memory operation (Article 10 data governance)
  • Transparency: All retrieval decisions are auditable — vector similarity, keyword matching, graph traversal. No black-box LLM decisions.
  • Risk classification: Local retrieval is minimal risk. No AI system makes autonomous decisions.
  • Right to explanation: You can trace exactly why a memory was recalled using slm trace "query"

Mode C: Partial Compliance

Mode C sends data to a cloud LLM provider. This means:

  • Data leaves your device (transmitted to the provider's servers)
  • You need a Data Processing Agreement (DPA) with your provider
  • The cloud provider's compliance status affects your overall compliance
  • Audit logs show which data was sent and when

Recommendation: Use Mode A or B for EU AI Act-regulated environments. Use Mode C only where cloud AI is explicitly approved by your organization.

GDPR

Right to Erasure (Article 17)

Delete memories matching a query:

slm forget "query matching memories to delete"

This permanently removes all matching memories, graph connections, and metadata. Because data is stored locally, there are no cloud logs to purge — deletion is immediate and complete.

To delete everything, remove the database:

rm ~/.superlocalmemory/memory.db

Right to Access (Article 15)

The database is a standard SQLite file at ~/.superlocalmemory/memory.db. You can copy it, query it directly with any SQLite tool, or use the dashboard to browse all stored data:

slm dashboard    # Visual browser at http://localhost:8765

Data Minimization (Article 5)

The entropy gate automatically filters out low-information messages during auto-capture. Only structured, high-value facts are stored — not raw conversation dumps.

Data Portability (Article 20)

The SQLite database is a standard, portable format. It can be copied to any machine and used immediately.

Access Control

Profile Isolation

Profiles provide complete data isolation:

slm profile create client-a
slm profile switch client-a

Memories in client-a are invisible to other profiles. There is no cross-profile data leakage.

Trust Scoring

Every agent that interacts with SuperLocalMemory has a Bayesian trust score (0.0 to 1.0):

  • Agents below the trust threshold are blocked from write and delete operations
  • Trust is updated based on outcome reports
  • View trust scores via the dashboard (Trust tab)

Tamper-Proof Audit Trail

All operations (store, recall, delete) are logged in a SHA-256 hash-chain audit trail. Each entry references the previous entry's hash — any tampering breaks the chain.

slm dashboard    # Compliance tab shows audit trail

HIPAA Considerations

SuperLocalMemory does not process Protected Health Information (PHI) by default. If you store PHI:

  • Use Mode A only (zero cloud)
  • Use profile isolation for patient contexts
  • Review audit logs regularly via dashboard

SuperLocalMemory does not provide BAA (Business Associate Agreement) coverage. Consult your compliance team before storing PHI.


Part of Qualixar | Created by Varun Pratap Bhardwaj

Clone this wiki locally