Personality Engine

Personality Engine — 6-System Behavior Framework

Goal: Make any AI agent feel alive — opinions, awareness, judgment, memory, timing sense, and engagement sensitivity. Works with trading agents, notification systems, personal assistants, or any proactive agent. Not just data delivery.

Architecture Overview

Trigger fires → engine.py orchestrator
              ↓
         selective_silence (should we stay silent?)
              ↓
         urgency_compute (how urgent is this 0.0-1.0?)
              ↓
         engagement_modifier (adjust for user response patterns)
              ↓
         variable_timing (schedule delivery based on urgency + time of day)
              ↓
         context_buffer (add back-references to earlier messages today)
              ↓
         editorial_voice (inject personality / opinions)
              ↓
         dedup (avoid repeats within rolling window)
              ↓
         send → iMessage (or other transport)

Plus two ambient systems:

• micro_initiations: Unprompted pings when conditions are met (quiet market, good streak, absence detected)
• response_tracker: Monitors engagement; adjusts urgency + suggests tuning

System 1: Editorial Voice — Opinion Injection

What: Each trigger type gets a personality — opinions that vary based on market state, portfolio P&L, signal confidence.

Per-trigger voice pools:

cross_platform (Kalshi vs Polymarket divergence)

• Bullish divergence (>5%): "Big divergence. One of these markets is wrong."
• Mild divergence (2-5%): "Mild divergence. Nothing screaming yet."
• Stale divergence (6+ hours old): "Divergence is stale — markets may have already repriced."

portfolio (user's holdings performance)

• +15% or better: "Good day. Portfolio's running."
• +5% to +15%: "Solid gains. Steady hand."
• -5% to +5%: "Flat day. Markets are grinding."
• -5% to -15%: "Rough patch. Check your stops."
• -15% or worse: "Heavy day. Buckle up for volatility."

x_signals (social signal scanner)

• Confidence ≥0.85 + matched position: "Strong signal. This feels real."
• Confidence 0.70-0.85: "New signal on [topic]. Worth watching."
• Confidence <0.70: "Noise signal. Low confidence."

edge (Kalshi edge detection)

• Edge >3%: "Fat edge. Worth a deep look."
• Edge 1-3%: "Mild edge. Keeping it on radar."
• Edge <1%: "Thin edge. Not worth the friction."

morning (daily brief)

• Monday: "New week. Here's the lay of the land."
• Friday: "Friday rundown. What matters before the close."
• Other: "Daily digest."

conflicts (overlapping triggers same hour)

• 2+ conflicts: "Tomorrow's a mess. Multiple overlaps."
• Lighter: "Heads up — couple things hitting together."

Customization point: Add trigger types by extending the VOICE_POOLS dict in editorial_voice.py. Each entry maps (trigger_name, market_state) → list of opinion strings.

System 2: Selective Silence — Knowing When NOT to Talk

What: Not every trigger fire deserves a message. Silent skips are explicit: "Skipped the brief — nothing worth your attention."

Content quality checks per trigger:

• morning_is_boring: If market vol <0.5%, no divergences, no edges → skip
• divergence_is_stale: If last message on this topic was <3 hours ago AND spread hasn't moved >0.5% → skip
• signals_are_noise: If all signals have confidence <0.65 AND no position matches → skip
• edge_is_weak: If all edges <1% → skip
• portfolio_is_flat: If daily P&L is -2% to +2% AND no major position changes → skip

Silence cadence:

• Max 1 silence message per day per user
• Only for expected triggers (morning, portfolio check, etc.)
• Never silence micro-initiations (those are the value)
• When silent, send explicit message: "Skipped the brief — nothing worth your attention."

Customization point: Adjust thresholds in selective_silence.py:

SILENCE_THRESHOLDS = {
    'vol_floor': 0.5,           # % vol threshold for morning silence
    'divergence_age_limit': 3,  # hours
    'signal_confidence_floor': 0.65,
    'edge_floor': 1.0,          # %
    'portfolio_flat_range': 2.0 # % P&L range
}

System 3: Variable Timing — Urgency Scoring + Time-of-Day Awareness

What: Schedule message delivery based on urgency (0.0-1.0) and time of day. A mild divergence at 6 AM gets sent immediately (threshold 0.9 before 7 AM). Same divergence at 10 PM gets held (threshold 0.35).

Per-trigger urgency base:

• cross_platform: (spread / 10%) * 0.6, capped at 1.0
  • 5% spread = 0.3 urgency
  • 10% spread = 0.6 urgency
  • 15%+ spread = 1.0 urgency
• portfolio: (abs(daily_pnl) / 10%) * 0.7, capped at 1.0
  • ±5% P&L = 0.35 urgency
  • ±15% P&L = 1.0 urgency
• x_signals: (confidence * 0.8) + (position_match ? 0.2 : 0), capped at 1.0
  • Confidence 0.85 + matched = 0.88 urgency
  • Confidence 0.70 + no match = 0.56 urgency
• edge: (edge_size / 5%) * 0.8, capped at 1.0
  • 2% edge = 0.32 urgency
  • 5% edge = 0.8 urgency
• meeting: (1.0 - minutes_away / 120) capped at 1.0
  • 30 min away = 0.75 urgency
  • 5 min away = 0.96 urgency

Time-of-day delivery thresholds:

• Before 7 AM: threshold 0.90 (almost everything gets sent)
• 7 AM - 9 AM: threshold 0.75 (morning crunch — moderate bar)
• 9 AM - 10 PM: threshold 0.45 (daytime — lower bar, let alerts through)
• 10 PM - 11 PM: threshold 0.35 (wind-down — only high urgency)
• 11 PM - 12 AM: threshold 0.85 (late night — back to high bar)
• 12 AM - 7 AM: threshold 0.90 (sleep time — very high bar)

Modifiers:

• Weekend: +0.10 urgency (weekends are boring, lower bar for engagement)
• Clustering prevention: If message sent <10 min ago, -0.20 urgency (space out messages)
• Daily fatigue: If 10+ messages today, +0.20 urgency threshold (user tired, fewer but higher-quality alerts)
• Random jitter: ±5% urgency (avoid machine-like precision)

Send logic:

adjusted_urgency = base_urgency * engagement_modifier ± jitter
if adjusted_urgency >= time_of_day_threshold:
    schedule_send(now or delayed based on urgency)
else:
    hold for next trigger

Customization point: Modify TIME_OF_DAY_THRESHOLDS and modifier constants in variable_timing.py:

TIME_OF_DAY_THRESHOLDS = {
    (0, 7): 0.90,    # midnight - 7 AM
    (7, 9): 0.75,    # 7 - 9 AM
    (9, 22): 0.45,   # 9 AM - 10 PM
    (22, 23): 0.35,  # 10 - 11 PM
    (23, 24): 0.85,  # 11 PM - midnight
}
MODIFIERS = {
    'weekend': 0.10,
    'clustering_prevention': 0.20,
    'daily_fatigue_step': 0.20,
}

System 4: Micro-Initiations — Ambient Awareness Pings

What: Unprompted messages when conditions are met. Not triggered by market events — triggered by meta-state.

Pools:

Cadence:

• Max 2 micro-initiations per week per user
• Skip on busy days (3+ regular alerts already sent that day)
• No repeats within 2 weeks (hash-based dedup: sha256(pool + date) in daily context)

US Holiday calendar (built-in awareness):

HOLIDAYS = {
    (1, 1): "New Year's Day",
    (1, 20): "MLK Day",
    (2, 17): "Presidents' Day",
    (3, 17): "St. Patrick's Day",
    (5, 26): "Memorial Day",
    (7, 4): "Independence Day",
    (9, 1): "Labor Day",
    (10, 13): "Columbus Day",
    (11, 11): "Veterans Day",
    (11, 27): "Thanksgiving",
    (12, 25): "Christmas",
}

Customization point: Add pools in micro_initiations.py:

MICRO_POOLS = {
    'QUIET_MARKET': {
        'condition': lambda ctx: ctx.vol < 0.3 and ctx.trade_count == 0,
        'messages': ["Quiet day. Markets are sleeping.", "No action today."],
    },
    'YOUR_POOL': {
        'condition': lambda ctx: your_logic_here(),
        'messages': ["Message 1", "Message 2"],
    }
}

System 5: Context Buffer — Daily Memory + Back-References

What: Messages can reference earlier messages from today. "That Kalshi/PM divergence I flagged at 9 AM widened to 15%." This makes the agent feel like it's thinking about past events, not just firing isolated alerts.

Per-trigger back-reference generation:

• cross_platform: If divergence was flagged earlier, compare current spread to earlier spread
• portfolio: If earlier portfolio message, show change since then
• x_signals: If same signal fired earlier, acknowledge the repeat with new data
• edge: Compare edge size to earlier edge on same market

Persistence: JSON file at ~/.openclaw/state/daily_context.json:

{
  "date": "2026-03-09",
  "messages": [
    {
      "time": "09:15",
      "trigger": "cross_platform",
      "spread": 7.2,
      "markets": ["kalshi", "polymarket"],
      "message_id": "msg_abc123"
    },
    {
      "time": "14:30",
      "trigger": "portfolio",
      "pnl": 12.5,
      "message_id": "msg_def456"
    }
  ],
  "silence_count": 1,
  "micro_count": 0,
  "sent_count": 5
}

Auto-reset: At midnight (UTC), clear context for fresh day.

Back-reference example:

Earlier (9:15 AM):  "Mild divergence. Kalshi 52%, Poly 48%."
Later (2:30 PM):    "That Kalshi/Poly spread I flagged this morning widened to 7% — now 55/48. Worth watching."

Customization point: Add back-reference logic for new trigger types in context_buffer.py:

def generate_backreference(trigger_type, current_data, history):
    if trigger_type == 'cross_platform':
        earlier = find_similar_trigger(history, 'cross_platform')
        if earlier:
            spread_change = current_data['spread'] - earlier['spread']
            return f"That spread I flagged {time_ago(earlier)} widened to {spread_change}%."
    return None

System 6: Response Tracker — Engagement Adaptation

What: Track user's response patterns. If user engages with 70%+ of messages, urgency stays high. If engagement <10%, adjust urgency down or suggest tuning.

Metrics per trigger type:

• sends: Count of messages sent
• engagements: Count of messages user responded to (within 1-hour window)
• ignores: Count of messages user didn't respond to
• avg_response_time: Average time from message to user response (minutes)

1-hour engagement window: If user responds to a message within 60 minutes, count it as engagement. After 60 min, assume ignored.

Urgency modifier:

• Engagement ≥70%: Multiply urgency by 1.3 (user likes these alerts, send more)
• Engagement 40-70%: Urgency × 1.0 (balanced)
• Engagement <10%: Urgency × 0.5 (user ignoring, tone it down)

Adaptation suggestion: After 10+ sends of a trigger type with <20% engagement, log:

⚠️ ADJUSTMENT SUGGESTION:
Trigger: x_signals
Sends: 12 | Engagement: 8% | Avg response time: Never

Consider:
→ Lower signal confidence floor (currently 0.65)
→ Reduce frequency (increase silence thresholds)
→ Check if message editorial voice is mismatched

Persistence: ~/.openclaw/state/response_tracker.json:

{
  "cross_platform": {
    "sends": 15,
    "engagements": 9,
    "ignores": 6,
    "avg_response_time": 23.5,
    "last_engagement": "2026-03-09T14:32:00Z"
  },
  "x_signals": {
    "sends": 12,
    "engagements": 1,
    "ignores": 11,
    "avg_response_time": null,
    "last_engagement": null
  }
}

Customization point: Adjust engagement thresholds and modifier multipliers in response_tracker.py:

ENGAGEMENT_THRESHOLDS = {
    'high': 0.70,    # ≥70% → 1.3x urgency
    'low': 0.10,     # <10% → 0.5x urgency
}
URGENCY_MULTIPLIERS = {
    'high': 1.3,
    'low': 0.5,
}
SUGGESTION_TRIGGERS = {
    'min_sends': 10,
    'max_engagement_for_suggestion': 0.20,
}

OpenClaw Ecosystem Integration

The Personality Engine works with any OpenClaw agent — it's domain-agnostic. Designed alongside the Prediction Market Trading Stack but applicable to any proactive agent that sends alerts, digests, or notifications.

Install the complete Prediction Market Trading Stack:

clawhub install kalshalyst kalshi-command-center polymarket-command-center prediction-market-arbiter xpulse portfolio-drift-monitor market-morning-brief personality-engine

Integration with OpenClaw Agents

Step 1: Import the engine

from personality_engine.engine import PersonalityEngine

engine = PersonalityEngine(user_id="user@example.com")

Step 2: Hook into proactive trigger system

In your agent's trigger handler (e.g., proactive_agent.py):

async def fire_trigger(trigger_type, data):
    # Your normal trigger logic
    message_content = generate_message(trigger_type, data)

    # Pass through personality engine
    should_send, scheduled_message = await engine.process_trigger(
        trigger_type=trigger_type,
        raw_message=message_content,
        market_data=data,
        urgency_context={'vol': market_vol, 'pnl': portfolio_pnl, ...}
    )

    if should_send:
        if scheduled_message.delayed:
            schedule_send(scheduled_message.content, delay=scheduled_message.delay_seconds)
        else:
            send_now(scheduled_message.content)
    else:
        log_silence_skip(trigger_type)

Step 3: Hook engagement tracker

When user responds to a message:

def handle_user_response(message_id, trigger_type, response_time_seconds):
    engine.response_tracker.log_engagement(trigger_type, response_time_seconds)

Step 4: Run micro-initiations

Add a separate cron job (every 30 min, low-overhead):

async def micro_initiations_check():
    micro_message = await engine.check_micro_initiations(context={
        'vol': market_vol,
        'trade_count': trades_today,
        'user_absence_hours': hours_since_last_engagement,
        'portfolio_streak': consecutive_days,
    })

    if micro_message:
        send_now(micro_message)

Step 5: Daily context reset

At midnight, engine auto-resets context. Manually trigger if needed:

engine.context_buffer.reset_daily()

Use Cases Beyond Trading

While the default voice pools and thresholds are tuned for prediction market trading, every system is designed for domain adaptation:

To adapt: swap the VOICE_POOLS dict in editorial_voice.py, update thresholds in selective_silence.py, and add domain-specific MICRO_POOLS in micro_initiations.py. See references/customization.md for full guide.

File Structure

personality-engine/
├── SKILL.md                           # This file
├── scripts/
│   ├── __init__.py
│   ├── engine.py                      # Main orchestrator
│   ├── editorial_voice.py             # Opinion injection
│   ├── selective_silence.py           # Content quality checks
│   ├── variable_timing.py             # Urgency + time-of-day
│   ├── micro_initiations.py           # Ambient pings
│   ├── context_buffer.py              # Daily memory
│   └── response_tracker.py            # Engagement tracking
├── references/
│   ├── systems-overview.md            # Architecture diagram + flow
│   └── customization.md               # Per-system customization guide
└── examples/
    └── integration-example.py         # Copy-paste integration template

Customization Quick-Start

Scenario 1: Your agent fires 50 alerts/day and user ignores most

1. Increase silence thresholds (selective_silence.py): Higher bars for what counts as "worth sending"
2. Lower variable_timing thresholds (variable_timing.py): Fewer messages slip through outside peak hours
3. Check editorial_voice: Opinions might not match user's style
4. Review response_tracker: Which trigger types have <20% engagement? Mute those.

Scenario 2: Your agent never initiates, only reacts

1. Customize micro_initiations pools: Add domain-specific conditions

   'LOW_VOLATILITY_OPPORTUNITY': {
       'condition': lambda ctx: ctx.vol < 0.2 and ctx.last_edge_size > 2.0,
       'messages': ["Calm market, good edge conditions. Might be time to scout."],
   }
   

2. Tune MICRO_CADENCE: Increase max from 2/week to 3-4/week if user loves it

Scenario 3: Your agent's opinions feel generic

1. Open editorial_voice.py and expand voice pools per trigger
2. Add market-specific opinions:

   'VOICE_POOLS': {
       'cross_platform': {
           'big_divergence': [
               "Big divergence. One of these markets is wrong.",
               "Spreads are blown out. Arb opportunity.",
               "Thick divergence — reality check time.",
           ],
           ...
   

3. Vary by user profile (trader vs long-term investor) by passing user_profile to engine init

Performance & State Management

State files (user's home directory):

• ~/.openclaw/state/daily_context.json (~5KB, resets daily)
• ~/.openclaw/state/response_tracker.json (~2KB, persistent)

Engine overhead:

• editorial_voice: <5ms (dict lookup)
• selective_silence: <10ms (threshold comparisons)
• variable_timing: <15ms (urgency calculation + time lookup)
• micro_initiations: <20ms (condition evaluation)
• context_buffer: <5ms (history lookup)
• response_tracker: <5ms (metrics lookup)

Total pipeline: ~60ms per trigger (negligible for async message delivery)

License

Part of the OpenClaw portfolio. Use freely in any agent.

Version History

v1.0.0 (2026-03-09)

• Initial 6-system framework
• Default configuration tuned for prediction market trading agents
• All systems designed for domain adaptation — swap voice pools, thresholds, and micro-initiation conditions for any use case