Memory Palace Skill

The Memory Palace is Optimus Pryme's long-term memory system. It stores, indexes,

and retrieves patterns, lessons, and knowledge accumulated over time, enabling the system to learn from experience and make increasingly intelligent decisions.

Core Capabilities

1. Historical Pattern Library

•Detect recurring patterns across campaigns and time
•Seasonal trend identification (Q4 surge, Prime Day, etc.)
•Day-of-week and time-of-day performance patterns
•Product lifecycle patterns
•Category-specific behaviors

2. Case-Based Reasoning

•Store successful strategies and their outcomes
•Learn from failures and near-misses
•Retrieve similar past scenarios for decision support
•"This worked last time" recommendations
•Context-aware pattern matching

3. Situational Memory

•Remember market conditions and responses
•Crisis response playbooks from experience
•Competitive action-reaction pairs
•Recovery strategies that worked

4. User Preference Learning

•Track which recommendations user accepts/rejects
•Learn risk tolerance and management style
•Adapt confidence thresholds per user
•Remember custom rules and constraints

5. Institutional Knowledge

•Vendor/supplier reliability tracking
•Seasonality calendars
•Product bundling success rates
•Keyword performance history

Pattern Types

Seasonal Patterns

json

{
  "pattern_type": "seasonal",
  "pattern_signature": {
    "month": "December",
    "category": "electronics",
    "metric": "sales_velocity"
  },
  "observed_effect": "3.2x increase in sales",
  "occurrences": 3,
  "success_rate": 1.0,
  "recommendation": "Increase bids 30-50% in November, scale budget by 3x"
}

Situational Patterns

json

{
  "pattern_type": "situational",
  "scenario": "high_acos_sudden_spike",
  "past_actions": [
    "decrease_bid_20_percent",
    "pause_underperforming_keywords"
  ],
  "outcome": {
    "acos_recovered": true,
    "recovery_time_days": 3,
    "revenue_impact": -5
  },
  "lessons": "Quick bid reduction more effective than keyword pausing"
}

User Preference Patterns

json

{
  "pattern_type": "user_preference",
  "preference_category": "risk_tolerance",
  "observed_behavior": {
    "auto_approves_below_budget": 100,
    "manual_review_above": 100,
    "rejects_aggressive_strategies": 0.8
  },
  "inferred_preference": "conservative",
  "confidence": 0.92
}

API Operations

Store Pattern

json

{
  "action": "store_pattern",
  "pattern": {
    "type": "seasonal",
    "signature": {...},
    "context": {...},
    "outcome": {...}
  }
}

Retrieve Similar Patterns

json

{
  "action": "find_similar",
  "current_situation": {
    "campaign_id": 123,
    "metrics": {...},
    "context": "high_acos"
  },
  "limit": 5
}

Response:

json

{
  "similar_cases": [
    {
      "similarity_score": 0.89,
      "past_scenario": {...},
      "actions_taken": [...],
      "outcome": {...},
      "recommendation": "Based on 3 similar cases, decrease bid by 15-20%"
    }
  ]
}

Learn User Preference

json

{
  "action": "update_preference",
  "user_action": "rejected",
  "recommendation_context": {
    "strategy": "aggressive",
    "budget_increase": 200
  }
}

Usage Patterns

Pattern 1: "What Worked Last Time?"

Scenario: Facing a performance issue

code

CURRENT: High ACoS on campaign X
MEMORY PALACE:
1. Search for similar past scenarios
2. Find 2 cases with high ACoS
3. Review actions taken and outcomes
4. Recommend best-performing approach

Pattern 2: Seasonal Prediction

Scenario: Approaching known seasonal event

code

TRIGGER: 6 weeks before Prime Day
MEMORY PALACE:
1. Retrieve Prime Day patterns from last 2 years
2. Identify: 3x demand spike, 40% higher competition
3. Recommend: Scale budget early, increase bids 2 weeks before
4. Alert user proactively

Pattern 3: User Preference Adaptation

Scenario: Generating recommendations

code

BEFORE RECOMMENDATION:
1. Check user's historical acceptance rate for similar recommendations
2. If user typically rejects "aggressive" → soften recommendation
3. If user auto-approves below $X → don't require manual approval
4. Adapt confidence thresholds to user's style

Database Schema

sql

-- From server/updates/04_meta_skills_tables.sql

memory_patterns (
  pattern_type,        -- 'seasonal', 'situational', 'user_preference'
  pattern_signature,   -- JSON description of pattern
  occurrences,         -- How many times observed
  success_rate,        -- Success rate (0.0 - 1.0)
  context,             -- Additional context
  first_seen,
  last_seen
)

case_library (
  scenario_description,
  actions_taken,      -- What was done
  outcome,            -- What happened
  lessons_learned,    -- Key insights
  created_at
)

Integration with Other Skills

Works with:

•orchestrator-maestro: Provide historical workflow success rates
•evolution-engine: Supply failure patterns to avoid
•consciousness-engine: Share prediction accuracy history
•grok-admaster-operator: Inform optimization decisions

Files

code

.agent/skills/memory-palace/
├── SKILL.md
├── scripts/
│   ├── pattern_miner.py          # Detect patterns in historical data
│   ├── memory_indexer.py         # Efficient storage and retrieval
│   └── similarity_matcher.py     # Find similar past scenarios
├── resources/
│   └── seasonal_patterns.json    # Known seasonal trends
└── tests/
    └── test_pattern_miner.py

Example Invocation

code

USER: "Should I increase my bid on this keyword?"

MEMORY PALACE ACTION:
1. Check: Has this keyword been optimized before?
2. Find: Yes, 3 months ago, bid increased 20% → ACoS worsened
3. Find: Similar keywords in this category respond better to 10% increases
4. Retrieve: User typically prefers conservative changes
5. Recommend: "Increase by 10% (historical data shows this range works better for this category)"

Notes

•Patterns strengthen with repeated observations
•Old patterns can "decay" if not recently observed
•User preferences are continuously refined
•All pattern storage respects privacy and data retention policies
•Patterns from different accounts are kept separate

This skill transforms Optimus Pryme from reactive to predictive, learning from every action and getting smarter over time.