Memory Amygdala

😱 Amygdala — Emotional Valence

Package: com.spectrayan.spector.memory.amygdala

Biological Analog: The amygdala is the brain's emotional processor. It assigns emotional significance to experiences — fear, joy, anger, relief — which profoundly influences how memories are encoded, stored, and retrieved. Emotionally charged memories are remembered more vividly and last longer.

---

The Concept

Every memory in Spector carries a valence score — a single byte (-128 to +127) representing its emotional coloring:

Range	Meaning	Examples
-128 to -50	Strongly negative	Critical errors, data loss, security breaches
-50 to -10	Mildly negative	Warnings, slow performance, minor bugs
-10 to +10	Neutral	Factual information, routine operations
+10 to +50	Mildly positive	Successful deployments, optimizations
+50 to +127	Strongly positive	Major breakthroughs, user praise, goals achieved

---

ValenceTracker

The ValenceTracker manages emotional coloring of memories:

public final class ValenceTracker {
    
    /**
     * Computes valence from text content analysis.
     * Uses keyword-based sentiment detection with configurable weights.
     */
    public byte computeValence(String text, MemorySource source) {
        float score = 0f;
        
        // Source-based bias
        if (source == MemorySource.PROCEDURAL) score += 0.1f;  // Rules are slightly positive
        if (source == MemorySource.OBSERVED && containsError(text)) score -= 0.5f;
        
        // Content-based sentiment
        score += sentimentScore(text);
        
        // Clamp and convert to byte range
        return (byte) Math.max(-128, Math.min(127, (int)(score * 127)));
    }
}

---

Valence-Filtered Recall

The most powerful use of valence is in recall filtering. The RecallOptions builder supports valence range filtering:

// Recall only negative-outcome memories (for debugging)
List<CognitiveResult> errors = memory.recall("database connection",
    RecallOptions.builder()
        .topK(10)
        .maxValence((byte) -10)     // Only negative memories
        .build());

// Recall only positive outcomes (for best practices)
List<CognitiveResult> successes = memory.recall("deployment strategy",
    RecallOptions.builder()
        .topK(5)
        .minValence((byte) 10)      // Only positive memories
        .build());

Phase 3 — Valence Filter in CognitiveScorer

Valence filtering happens at Phase 3 of the scoring pipeline — before the expensive SIMD vector math:

// Phase 3: Valence Filter (~2 cycles)
byte valence = segment.get(LAYOUT_VALENCE, offset + OFFSET_VALENCE);
if (valence < minValence || valence > maxValence) continue;

Cost: 2 CPU cycles — a single byte read and two comparisons. Records outside the valence range are eliminated before Phase 5's ~200-cycle SIMD computation.

---

Use Cases

1. Debugging: "What Went Wrong?"

An agent can filter for negative-valence memories when debugging:

// "Show me only memories associated with failures"
memory.recall("connection timeout",
    RecallOptions.builder()
        .maxValence((byte) -10)
        .synapticFilter("database", "error")
        .build());

2. Best Practices: "What Worked Well?"

// "Show me successful approaches"
memory.recall("deployment strategy",
    RecallOptions.builder()
        .minValence((byte) 10)
        .synapticFilter("deployment")
        .build());

3. Balanced Recall: Full Emotional Range

By default, no valence filter is applied — the agent sees the full emotional spectrum. The valence still influences recall indirectly because the FlashbulbPolicy pins emotionally intense memories at higher importance.

---

Storage

Valence is stored in the 32-byte synaptic header at offset 30 as a single signed byte:

Offset 30: [1B valence] — signed byte [-128 to +127]

This costs exactly 1 byte per memory — negligible overhead for a powerful filtering dimension.

---

Next Steps

• :material-link: [[Hebbian — Association Learning|Memory--Hebbian]] — "neurons that fire together wire together"
• :material-head-cog: [[Dopamine — Surprise Detection|Memory--Dopamine]] — auto-importance scoring
• :material-lightning-bolt: [[6-Phase Scoring Pipeline|Memory--Scoring-Pipeline]] — where valence filtering happens