Framework

Privacy

Personal Data: Local Infrastructure

Privacy Policy — Local Infrastructure Platform Use of My Data

Audience: Mark Myers (internal reference)

Purpose: Documents how the Local Infrastructure platform handles my data on this system

Effective Date: March 2, 2026

Governing Document: CONSTITUTION.md ("How I Work With AI," v5)

What This Covers

This document describes what data the Local Infrastructure platform accesses, stores, transmits, and retains on the Mac Mini where Golgi operates. It is a self-audit, not a legal compliance document. It exists so I know exactly what my agent touches and where my data goes.

What Local Infrastructure Accesses on This Machine

Local Infrastructure operates under a single-user personal assistant trust model. The agent has broad access to the local system within the scope I've configured:

File system access:

  • Full read/write access to ~/.local-infra/workspace/ (governance files, memory, project files)
  • Full read/write to ~/.local-infra/ (configuration, credentials, session transcripts, logs)
  • Shell access via tool use (can execute commands, read/write files, run scripts)
  • Access to any file or directory the OS user account can reach

    • Messaging channels:

  • Telegram messages sent to/from the bot account (processed in real time)
  • SMS messages routed through Twilio webhook (queued in Firebase, relayed via Telegram)
  • Potential access to Discord, WhatsApp, or other channels if configured

    • Network access:

  • Web browsing and search via tool use
  • API calls to external services (GitHub, Vercel, Firebase, etc.)
  • Outbound HTTP requests as needed for tool execution

    • What Data Flows to Anthropic's API

    Every interaction with Golgi sends data to Anthropic's Claude API for processing. This includes:

  • Message content: The text of my messages and Golgi's responses
  • System prompts: Workspace governance files loaded at session start (CONSTITUTION.md, SOUL.md, USER.md, AGENTS.md, etc.)
  • Tool use context: Commands executed, file contents read, search results — anything that enters the conversation context window
  • Session history: Prior messages in the current session (conversation context)

    • What Anthropic does with API data (as of March 2026):

  • API data is retained for 7 days (reduced from 30 days as of September 2025), then automatically deleted
  • API data is not used for model training under commercial/API terms
  • Organizations can opt into 30-day retention for audit purposes via Data Processing Addendum
  • Flagged content may be retained up to 2 years (inputs/outputs) or 7 years (trust & safety scores) if a safety policy violation is detected

    • What this means in practice: Every governance file, every memory entry, every message I send through Golgi passes through Anthropic's infrastructure. It's processed, held briefly, and deleted. But during that window, it exists on their servers.

    What Data Is Stored Locally vs. Transmitted

    Data Type
    Stored Locally
    Transmitted Externally
    Governance files (CONSTITUTION.md, SOUL.md, etc.)
    Yes — ~/.local-infra/workspace/
    Yes — loaded into API context each session
    Memory files (MEMORY.md, daily notes)
    Yes — ~/.local-infra/workspace/memory/
    Yes — loaded into API context when read
    Session transcripts
    Yes — ~/.local-infra/agents/*/sessions/*.jsonl
    No — stored locally only
    Configuration and credentials
    Yes — ~/.local-infra/local-infra.json, credentials/
    No — used locally for auth
    Telegram messages
    Processed in transit
    Yes — Telegram servers + Anthropic API
    SMS messages
    Queued in Firebase
    Yes — Twilio + Firebase + Anthropic API
    Shell command output
    Transient (in context window)
    Yes — enters API context if part of conversation
    Screenshots and images
    Yes — ~/.local-infra/workspace/screenshots/
    Only if explicitly sent via Telegram or entered into context
    Gateway logs
    Yes — ~/.local-infra/logs/
    No — local only

    Token and Credential Handling

    Credentials are stored locally under ~/.local-infra/ in the following locations:

  • Gateway auth token: local-infra.json or environment variable
  • Anthropic API key: Agent auth profile
  • Telegram bot token: Channel config or environment variable
  • Firebase/Twilio credentials: Environment variables or config files
  • GitHub tokens: Environment variables

    • Local Infrastructure does not provide built-in encryption for stored credentials. Protection relies on:

  • OS-level file permissions (directories at 700, files at 600)
  • Full-disk encryption on the host (FileVault on macOS)
  • Environment variables preferred over config file storage where possible

    • The security hardening completed in late February 2026 addressed file permissions, credential isolation, and gateway authentication.

    Memory and Persistence

    Golgi maintains persistent state across sessions through:

  • MEMORY.md — Curated long-term memory, updated by the agent during main sessions
  • Daily memory files — memory/YYYY-MM-DD.md, raw logs of session activity
  • Session transcripts — Full conversation logs stored as JSONL files
  • SQLite databases — Performance tracking (e.g., golgi_performance.db)
  • Heartbeat state — JSON file tracking periodic check-in status

    • All persistent data is stored locally on the Mac Mini. Session transcripts may contain any data that entered the conversation — including file contents, personal messages, command output, and API responses.

    What I Can Delete, Export, or Audit at Any Time

    Full control over all local data:

  • Delete any file under ~/.local-infra/ at any time
  • Export workspace files, memory, and transcripts by copying them
  • Audit session transcripts to see exactly what was sent and received
  • Review gateway logs for operational history
  • Git history (as of March 2, 2026) provides version tracking for governance files

    • API-side data:

  • Cannot delete data during Anthropic's 7-day retention window (it auto-deletes)
  • Can review Anthropic's data handling via their Privacy Center (https://privacy.claude.com)
  • Not opted into extended retention or training data contribution

    • Local Infrastructure's Own Privacy Posture

    Local Infrastructure is an open-source, self-hosted platform. There is no centralized Local Infrastructure service collecting data from users. As noted in their security documentation: the platform operates on a personal assistant trust model where the operator (me) controls the trust boundary.

    Local Infrastructure does not have a traditional privacy policy because it doesn't collect user data — it runs on my hardware, under my control. The privacy implications are defined by how I configure it and which external services I connect.

    Risks I've Accepted

    By running an autonomous agent with shell access, I've accepted the following:

  • Broad local access: Golgi can read/write any file my user account can reach. A misconfigured or compromised agent could access sensitive data beyond the workspace.
  • API data exposure: Every conversation passes through Anthropic's infrastructure. Even with 7-day deletion, data exists on external servers temporarily.
  • Credential proximity: API keys, tokens, and credentials are stored on the same machine the agent operates on. Defense relies on file permissions and OS-level encryption.
  • Messaging channel exposure: Messages processed through Telegram and SMS pass through third-party infrastructure (Telegram servers, Twilio, Firebase).
  • Session transcript accumulation: Conversation logs grow over time and may contain sensitive data from past sessions if not pruned.
  • Agent autonomy risk: Golgi can execute shell commands, write files, and take actions without per-action approval. The governance framework (CONSTITUTION.md, SOUL.md) constrains this, but the technical capability is broad.

    • Institutional Data Boundary

    No student records, protected educational data, or institutionally governed data enters this system. This is a hard constraint established in CONSTITUTION.md. FERPA-protected information, Augusta University student data, and any information governed by institutional data policies are never processed through Local Infrastructure, Golgi, or any connected service.

    This system processes Mark's personal and professional communications, drafts, and operational data only.

    *This document is part of the privacy policy suite. See also: PERSONAL_DATA_AGENTIC.md, PERSONAL_DATA_EXTERNAL.md, PUBLIC_USER_PRIVACY_POLICY.md.*

    *Governing document: CONSTITUTION.md ("How I Work With AI," v5)*