Skip to content

Privacy Model & PrivacyShield Engine

Neural Memory handles sensitive personal context. Its privacy architecture is built around strict local-first isolation, proactive client-side redaction, and granular user consent.

1. Core Privacy Guarantees

  • Zero Outbound Data by Default:
  • In base mode, no data ever leaves your computer. The API binds strictly to 127.0.0.1.
  • In Standalone mode, all data is stored inside a local SQLite property graph at: ~/Library/Application Support/NeuralMemory/memory.db.
  • Even when LLM enrichment is enabled (LLM_ENABLED=true), all sensitive credentials and identifiers are redacted locally before transmission.
  • Granular Independent Opt-Ins:
  • Activity monitoring (app/window focus), screenshot capture, and keystroke buffer capture require separate, explicit toggles in Settings.
  • On a fresh installation, all capture engines start disabled.
  • Ephemeral Data Pruning (Dream Mode):
  • Raw keyboard, mouse, and window events are automatically pruned after 48 hours (configurable retention period).
  • Only durable semantic knowledge nodes (Decision, Commitment, Meeting, Reflection) persist in long-term memory.

2. PrivacyShield Redaction Pipeline

Every interaction bundle ingested by Neural Memory is processed by the local PrivacyShield service (server/src/kg_mcp/services/privacy_shield.py) before storage or LLM vision processing.

2.1 Application & Window Deny-List

Neural Memory instantly drops and suppresses any capture originating from sensitive contexts: - Password Managers & Vaults: 1Password, Bitwarden, Apple Keychain Access, KeePassXC, Enpass. - Private Browsing: Window titles containing "Incognito", "InPrivate", or "Private Browsing". - Financial Applications: Banking portals, payment processors, and crypto wallets.

2.2 Proactive In-Memory Redaction

Typed text and contextual strings are scrubbed via dedicated heuristic and cryptographic validators:

Sensitive Pattern Validation Method Replacement Mask
Credit Card Numbers Regex pattern matching + Luhn Algorithm mathematical checksum validation [REDACTED_CREDIT_CARD]
IBAN Numbers ISO 13616 international standard validation [REDACTED_IBAN]
OpenAI API Keys sk-[a-zA-Z0-9]{32,64} [REDACTED_API_KEY]
Google Gemini API Keys AIza[0-9A-Za-z-_]{30,45} [REDACTED_API_KEY]
GitHub Tokens gh[pousr]_[0-9A-Za-z]{36} [REDACTED_API_KEY]
Generic Bearer Tokens Bearer [A-Za-z0-9_-]{20,} Bearer [REDACTED]
Passwords / PINs Contextual field heuristics (password=..., pin=...) [REDACTED_CREDENTIAL]

3. Storage & Cryptographic Tokens

  • Loopback Isolation: The HTTP API binds only to 127.0.0.1:8765. It cannot be reached across local network subnets or Wi-Fi.
  • Bearer Token: Access to all mutation and private query endpoints requires a 256-bit cryptographically secure token.
  • In Standalone mode: Persisted in ~/Library/Application Support/NeuralMemory/token.txt with 0600 permissions.
  • In macOS UI app: Stored in the secure system macOS Keychain.
  • Screenshot Transient Storage: Raw screenshot bitmap pixels are never persisted into the graph database. They are processed in memory for perceptual visual hashing (dhash) to eliminate duplicates and immediately discarded.

4. User Controls & Best Practices

  1. Pause / Private Mode: Available directly from the macOS Menu Bar. One click pauses all contextual observation while you handle personal matters.
  2. Database Wipe:
  3. To completely erase the standalone database: 
    rm -f ~/Library/Application\ Support/NeuralMemory/memory.db*
  4. To erase the Docker database: 
    docker compose down -v
  5. Never Share Your Token: Treat token.txt and .env as sensitive credentials. Do not commit them to version control.