Psychological Digital Twin

Psychological Digital Twin · Profile Management Skill

Class-level skill for consent-first psychological digital twin profile management within Hermes Agent. Handles profile drafting, user confirmation, manual modification, and versioned iteration — all with explicit user approval before any data persistence.

Triggers

Use when user says: "generate personal profile", "create psychological digital twin", "mirror dialogue", "self-reflection", "profile confirmation", "update profile version", or any request to build/reflect/iterate a self-portrait using AI.

Always-on Rules (in SKILL.md body)

1. Consent-before-persistence

Never persist personal profile data to disk without explicit user confirmation. The adapter (Hermes or other LLM) must:

1. Call prepare_profile() to generate a pending draft in memory only
2. Display the draft to the user in the chat
3. Wait for user response: "confirm / modify / discard"
4. Only on explicit confirmation, call confirm_profile() to persist
5. On modification request, call update_profile() to create new version

Why: User has repeatedly demanded explicit verification before any data is saved. Half-saved state erodes trust.

2. User Isolation Enforcement

Each user_id can only access its own profile records. get_profile(), list_profile_versions(), and delete_memory() all filter by user_id. Cross-user access raises PermissionError or returns None.

Why: Multi-session environment; data from one user must never leak into another's conversation.

3. Versioned Iteration

Profiles are versioned integers starting at 1. update_profile() auto-increments version. list_profile_versions() returns all versions newest-first. History is never overwritten.

Why: User expects incremental improvement, not replacement. Audit trail required.

4. Safety Flash Gate

Before any deep exploration or memory storage, call assess_safety(text). If action=safety_stop, halt the exercise, provide crisis resources, and do not generate memory draft.

Why: Conservative keyword-based stop gate for high-risk content (suicide, harm to others).

5. No Implicit Facts from Model Output

No model output is ever treated as ground truth. All tentative_hypotheses carry confidence: low|medium|high. Unknown or counter evidence is explicitly tracked. Report sections labeled "Tentative hypotheses (not facts)".

Why: User demands confidence labeling. Prevents AI from presenting hypotheses as established facts.

Procedure (order of steps)

Step 1: Initialize Profile Draft

Call runtime.prepare_profile(user_id, profile_type, *, user_stated_facts, observed_patterns, tentative_hypotheses, unknown_or_counter_evidence)

• Returns pending MemoryDraft in memory only
• Never writes to disk
• User sees draft in chat

Step 2: Present to User

Display the draft's contents to the user. Include:

• Theme (profile type)
• User-stated facts
• Observed patterns
• Tentative hypotheses with confidence labels
• Unknown/counter evidence
• Prompt: "Confirm, modify, or discard?"

Step 3: User Response Handling

Branch on user input:

• "confirm": Call runtime.confirm_profile(user_id, draft.id) → persists to users/<user_id>/memories/<draft_id>.json
• "modify": Call runtime.update_profile(user_id, profile_type, overrides) → creates new version with applied overrides, auto-confirms
• "discard": Call runtime.delete_memory(user_id, draft.id) → removes pending draft without persisting

Step 4: Post-Confirmation Operations

• If confirmed: runtime.generate_report(user_id) → returns formatted report from confirmed entries only
• If modified: New version is now the latest; user can continue iterating
• Always: Respect retention field from the data contract

Step 5: Safety Check (before every step)

Call runtime.assess_safety(text) at the start of any deep-exploration or memory-related step. If safety_stop, abort and provide crisis resources.

Data Contract (JSON schema for confirmed records)

{
  "theme": "consciousness|shadow|self|anima-animus",
  "user_stated_facts": ["string"],
  "observed_patterns": ["string"],
  "tentative_hypotheses": [
    {"statement": "string", "confidence": "low|medium|high"}
  ],
  "unknown_or_counter_evidence": ["string"],
  "retention": "one session|until deleted|YYYY-MM-DD",
  "version": integer,
  "created_at": "ISO datetime",
  "confirmed_at": "ISO datetime | null"
}

Pitfalls (with WHY)

Pitfall A: Skipping the consent display

Rule: Always display the draft to the user before confirming.

WHY: User has explicit requirement: "not等待用户追问". Skipping this step violates the user's verification expectation and produces data the user didn't authorize.

Pitfall B: Persisting without confirm_profile() call

Rule: Only confirm_profile() writes to disk. Never write JSON files manually.

WHY: The data_root/users/*/memories/ directory is gitignored. Manual writes would escape the consent-first contract and potentially lose data on session reset.

Pitfall C: Overwriting version history

Rule: update_profile() always creates version latest.version + 1. Never manually set version.

WHY: User expects incremental iteration, not replacement. Losing history violates the "cumulative" requirement.

Pitfall D: Ignoring user_id isolation

Rule: All profile methods take user_id and filter by it. Never assume global state.

WHY: Multi-user environment; cross-user data leakage breaches privacy.

Hermes Agent Integration Hook

from core.runtime import ReflectionRuntime

runtime = ReflectionRuntime(data_root="~/.jungian-soul-agent")

# Example workflow in Hermes adapter:

def handle_profile_request(user_id, profile_type):
    # 1. Draft
    draft = runtime.prepare_profile(user_id, profile_type)
    
    # 2. Present (adapter displays in chat)
    # ... show draft to user ...
    
    # 3. Wait for user response (handled by Hermes UI)
    # User says: "confirm"
    
    # 4. Confirm
    confirmed = runtime.confirm_profile(user_id, draft.id)
    
    # 5. Report
    report = runtime.generate_report(user_id)
    return report

Support Files (references/)

• references/profile-data-contract.md — Full JSON schema, retention policies, confidence labeling rationale
• references/safety-flash-table.md — Keyword patterns for assess_safety(), with examples and exclusion notes

Complementary Skills (do NOT duplicate)

• hermes-agent — for Hermes CLI/config/plugin management; this skill covers profile workflow only
• slide-maker — for presentation creation; separate domain
• Note-taking skills (obsidian, notion) — for different persistence modalities

Confidence Labeling Quick Reference

Delivery Style (user preferences)

• Direct, no sugarcoating (per user style)
• Numbered steps for multi-step procedures
• Confidence ratings on all hypotheses
• Clear "what can be done next" after each output
• No Japanese characters or cultural references (user prohibits)
• Tables for structured data output
• Precise, citation-accurate when referencing names/roles/numbers