hermes - 💡(How to fix) Fix Shared user identity layer - auto-collected from conversation, shared across all profiles

Problem

Hermes currently has no shared identity layer. User information (preferences, tools, workflow, contacts) is learned separately by each profile through memory, and there is no mechanism to share this learning across profiles or persist it in a structured format.

Current behavior:

• Each profile maintains its own memory of the user
• What default learns about the user is invisible to xiaoming
• No structured schema for user identity
• No mechanism to enrich identity over time from conversation

Proposed Solution

Add a lightweight identity layer at ~/.hermes/identity.yaml that:

1. Is read at session start by every profile
2. Is written by agents during conversation (with user confirmation)
3. Contains structured fields: name, preferences, tools, workflow, contacts

Design Principles

1. Setup asks nothing new - just name (already exists)
2. Everything else is auto-collected from conversation
3. Agent detects new information and asks user before writing
4. One file, all profiles share it

Auto-Collection Flow

During conversation, the agent observes:

• User mentions a tool: I just installed Docker

  • Agent: I see you use Docker. Want me to add it to your identity?
  • User: Sure
  • Agent writes to identity.yaml

• User expresses a preference: Dont ask me yes/no questions, just do it

  • Agent: Got it. Ill add that to your communication preferences.
  • Agent writes to identity.yaml

• User mentions a project: Im working on a digital native project

  • Agent: Want to add this to your project list?
  • User: Yes
  • Agent writes to identity.yaml

File Format

File at ~/.hermes/identity.yaml. Structure:

identity_version: 1.0
profile:
  name: user
  preferred_name: boss
communication:
  style: concise Chinese, direct
  avoid:
    - asking yes/no questions
    - long explanations
tools:
  - aider
  - docker
projects:
  - name: digital native
    description: world model project
changelog:
  - date: 2026-05-30
    author: auto-discovery
    change: added tools.aider

Key Behaviors

1. Auto-loaded at session start - injected into system prompt, no skill needed
2. Shared across profiles - default, xiaoming, model-trainer all read the same file
3. Agent-writable - with user permission, agent updates the file
4. Immediate effect - updated fields are visible in the next turn
5. CLI - hermes identity for manual view/edit

What Gets Auto-Collected

• Tools the user mentions (docker, aider, vscode, etc.)
• Communication preferences (style, things to avoid)
• Workflow rules (always write plan first, red lines)
• Active projects
• Contact methods (when mentioned)
• Service configurations (proxy, cloud providers)

What Never Gets Auto-Collected

• Secrets (passwords, API keys, tokens)
• Temporary state
• One-off opinions
• Anything the user declined

Benefits

• One conversation trains every profile
• No setup burden - just talk naturally
• Identity improves over time without manual configuration