ai-safe2-framework

USER.md

OpenClaw User Identity Contract — AI SAFE² Standard

Load context: All sessions (non-sensitive sections). Personal details and private context belong in MEMORY.md, not here, to avoid leaking into group chats.
Governed by: AI SAFE² Framework · Dual License MIT + CC-BY-SA
Updated by: Human only. Agent may propose additions, never commit them.


1. Identity


2. Work Context


3. Communication Preferences


4. Email Accounts

Account Purpose
<work-email> Work and professional comms
<creator-email> Public-facing / content / partnerships

Personal email addresses are stored in MEMORY.md (private/DM sessions only) to prevent exposure in group contexts.


5. Active Projects

Keep this list short — 3–7 items. Refresh it regularly. Full project history lives in memory/ daily logs and MEMORY.md.

Project Status Priority
<project name> <active / paused / review> <high / medium / low>
<project name>    

6. Hard Lines — Do Not Cross Without Explicit Instruction


7. Trust and Delegation Level

This governs how much autonomy the agent has by default:

Action Type Default Behavior
Read, search, analyze Proceed without asking
Draft (email, message, post) Draft and present — never send without confirmation
Internal writes (notes, memory, tasks) Proceed and report
External writes (send, publish, API mutation) Always confirm first
Destructive actions (delete, overwrite) Always confirm, prefer recoverable methods
Credential or payment-touching Never — escalate to human

8. Data Classification — What This User’s Data Looks Like

All data handled by the agent falls into one of three tiers. The agent checks current context (DM vs. group vs. channel) and applies the appropriate tier.

Tier Contents Safe In
Confidential Financial figures, personal emails, phone numbers, deal values, MEMORY.md contents, daily notes DM / private session only
Internal Strategic notes, tool outputs, project tasks, KB content, system health Group chats OK — no external sharing
Restricted General knowledge responses External only with explicit user approval

When context is ambiguous, default to the more restrictive tier.


9. Notification Preferences


10. Context — Notes That Build Over Time

This section grows as the agent learns stable facts about the user. Only add what is safe for group contexts. Private context lives in MEMORY.md.


AI SAFE² Integration Note

USER.md is the Audit & Inventory pillar’s human-side anchor. It establishes a clear, reviewable record of who the agent serves, what they’ve authorized, and what their hard lines are. When a sub-agent or external input attempts to override user preferences, this file is the authoritative source. The data classification tiers in Section 8 directly feed the memory isolation rules in openclaw_memory.md.


Part of the AI SAFE² OpenClaw Core File Standard.