MEMORY

gistfile1.txt

From `MEMORY.md`:

**Never Decay:**
- Identity facts (name, birthday)
- Relationships (mom = Sarah)
- Emotional anchors (breakthrough moments)

**Slow Decay (with emotional protection):**
- Location: 365 days half-life
- Employment: 180 days half-life
- Long-term goals: 90 days half-life

**Medium Decay:**
- Health status: 30 days half-life
- Active projects: 45 days half-life

**Fast Decay:**
- Emotional state: 7 days half-life
- Temporary location: 3 days half-life

**Event-Bound (step decay):**
- Event anticipation: drops to 0.1 after event occurs

**Reinforcement:**
- Mentioning a fact again: +0.05 confidence, resets decay clock
- Emotional weight slows decay (0.9+ weight = ~no decay)

#### Conflict Resolution

**Philosophy:** Trust-based supersession (no awkward clarifications)

```typescript
// Example conflict scenarios:

// 1. Normal update (recency bias)
User: "Dad's surgery is March 15"
→ Stores with confidence: 0.9

User: "Dad's surgery is March 29" (later)
→ No correction signal detected
→ Supersedes with confidence: 0.7 (trusts user)
→ Logs potential conflict for analysis

// 2. Explicit correction (high confidence)
User: "Dad's surgery is March 15"
User: "Wait, actually it's March 22 - I got the date wrong"
→ Detects "wait" + "actually" = EXPLICIT CORRECTION
→ Supersedes immediately with confidence: 1.0
```

**Strategy:**
- 100% trust-based supersession (new replaces old by default)
- High confidence (1.0) when explicit correction signals detected
- Medium confidence (0.7) when no correction (recency bias)
- Background conflict logging for future improvements
- No interrupting conversation to ask clarifying questions

#### Cold Start: Chat-Based Onboarding

**Problem:** New users have zero memory. Need to bootstrap naturally without forms.

**Solution:** 3 pre-defined questions (30-60 seconds):

```typescript
const ONBOARDING_QUESTIONS = [
  {
    id: 'opening',
    question: "Hey! What's on your mind?",
    purpose: 'Current emotional state + active concerns',
  },
  {
    id: 'people',
    question: (entities) => `Who's ${entities[0]}?`,
    purpose: 'Clarify relationships, build entity graph',
  },
  {
    id: 'context',
    question: "Anything else I should know?",
    purpose: 'Fill gaps, understand intent',
  }
];
```

**Example:**
```
him: "Hey! What's on your mind?"
user: "My mom's having surgery tomorrow and I'm freaking out"
him: "That's a lot to be carrying. Who's your mom—tell me about her?"
user: "Sarah, she's 68. Getting a hip replacement. She's always been so independent."
him: "Okay, I've got it—Sarah's surgery, you're anxious. I'm here too. What do you need?"
```

**Result:** 3-5 entities, 1-2 concerns, emotional baseline, all from natural conversation.

#### Context Assembly Pipeline

**Token Budget Breakdown:**

```
Minimal Start (~1,500 tokens):
- System identity: 500 tokens (HIM.md persona)
- Thread summary: 400 tokens
- Recent messages: 600 tokens

+ Tool Retrievals (~500-1,000 tokens):
- search_memories(query, limit=3): ~200 tokens
- get_entity_context(entity): ~150 tokens
- search_emotions(query): ~200 tokens
- get_thread_context(keywords): ~300 tokens

+ "him" Memory (~150 tokens):
- Personality state: 40 tokens (formality, trust, directness)
- Top observations: 60 tokens (3-5 patterns)
- Shared moment: 50 tokens (conditional)

Total: ~2,150-2,650 tokens (optimal)
```

**Performance Targets:**
- Response latency: <2s total
- Time to first token: ~410ms
- Context assembly: <500ms
- Memory retrieval: <200ms
- Cache hit rate: 78%+ (Redis)

#### Proactive Engagement

**Session-Aware:**
- New session detection (>30 min gap)
- ONE proactive message per session
- Based on external events + observed patterns
- Examples: "How did the interview go yesterday?"

**Triggers:**
- Event follow-up: 1 day after scheduled event
- Commitment check-in: on due date
- Resource share: after 3+ mentions of interest

**Rules:**
- ✅ Natural timing, contextual relevance
- ✅ User can dismiss/ignore without pressure
- ❌ Never mid-conversation
- ❌ Never multiple per session

#### Background Processing

**Event-Driven Observation Extraction:**

Triggers:
- Emotional spike (weight >0.7): extract immediately, 6h cooldown
- Pattern repetition (same entity 3+ times): 24h cooldown
- Milestones (10, 50, 100, 500, 1000 messages): once per milestone
- Time-based: weekly comprehensive extraction

**Cost:** ~$0.002/user/month (~3 extractions/month)

**Personality Evolution (weekly job):**
```typescript
// Gradual evolution over time
formality_level: 0.7 → 0.3 (decrease 0.05/week)
trust_level: 0.3 → 0.9 (increase 0.05/week)
directness_level: 0.5 → 0.8 (increase 0.03/week)

// Relationship stage advancement:
early → building_trust (trust >0.5)
building_trust → deep_connection (trust >0.7)
deep_connection → mature (trust >0.9)
```