Importance Fusion (ICNU)¶

The ICNU Importance Fusion system computes a memory's importance score at ingestion time by blending four signals: Interest, Challenge, Novelty, and Urgency — then applying salience profile topic boosts and persona-based modulation.

The Problem¶

Without ICNU, importance is determined solely by the Surprise Detector — a statistical outlier test based on how "surprising" a memory's embedding is relative to recent memories. This works well for detecting unusual information, but has blind spots:

• A memory about a user's urgent deadline might not be statistically surprising
• A memory about a challenging technical problem might have a common embedding
• A memory that the agent finds interesting has no way to signal that interest

ICNU adds three LLM-provided signals alongside the existing novelty signal, then layers salience profiles and persona traits on top to produce a richly personalized importance score.

The Full Importance Pipeline¶

flowchart LR
    ICNU["Stage 1\nICNU Fusion\nI·C·N·U → raw importance"] --> SALIENCE["Stage 2\nSalience Profile\ntopicBoost × raw importance"]
    SALIENCE --> PERSONA["Stage 2b\nPersona Modulation\nvalence bias + arousal sensitivity"]
    PERSONA --> FLASH{"Stage 3\nFlashbulb\ncheck"}
    FLASH -->|"z ≥ threshold"| PIN["Pin at 10.0"]
    FLASH -->|"normal"| FINAL["Final importance\n(written to header)"]

    style ICNU fill:#f39c12,color:white
    style SALIENCE fill:#0984e3,color:white
    style PERSONA fill:#7c3aed,color:white
    style PIN fill:#d63031,color:white
    style FINAL fill:#00b894,color:white

Stage 1: ICNU Formula¶

Where:

The weights \(w_i\) are configurable and auto-normalize to sum=1.0:

Output Range¶

The formula maps to rawImportance ∈ [0.05, 10.0]:

• 0.05 — All signals zero (routine, uninteresting, familiar, non-urgent)
• 10.0 — All signals maximal (interesting, challenging, novel, urgent)

Stage 2: Salience Profile Boost¶

After ICNU computes the raw importance, the salience profile applies a topic boost based on the user's declared interests and disinterests:

How Topic Boost Works¶

The salience profile contains interest topics with pre-computed embeddings. At ingestion time, cosine similarity is computed between the memory and each interest:

flowchart LR
    MEM["Memory embedding"] --> SIM["Cosine similarity\nwith each interest topic"]
    SIM --> MATCH{"similarity >\nthreshold (0.5)?"}
    MATCH -->|"Yes"| BOOST["Apply interest level\nmultiplier × similarity"]
    MATCH -->|"No"| PASS["No boost (1.0×)"]
    BOOST --> FINAL["topicBoost"]
    PASS --> FINAL

    style MEM fill:#4a6fa5,color:white
    style SIM fill:#0984e3,color:white
    style BOOST fill:#00b894,color:white
    style FINAL fill:#f39c12,color:white

Interest Levels¶

Example¶

Memory: "PostgreSQL query optimizer regression after upgrade"

Interest: "database performance" (CRITICAL, multiplier=2.0)
  cosine("database performance", memory) = 0.82  → above threshold
  topicBoost = 2.0 × 0.82 = 1.64

Disinterest: "meeting notes" (IGNORE, multiplier=0.1)
  cosine("meeting notes", memory) = 0.12  → below threshold
  → No dampening applied

Raw ICNU importance: 5.0
Boosted importance: 5.0 × 1.64 = 8.2

Stage 2b: Persona Modulation¶

If the salience profile includes a cognitive persona, three additional traits modulate the importance:

flowchart LR
    BOOSTED["Boosted importance\n(from salience)"] --> VB["Valence bias\nshift emotional baseline"]
    VB --> AS["Arousal sensitivity\nmodulate decay resistance"]
    AS --> SR["Self-relevance\nentity identity boost"]
    SR --> FINAL["Final importance"]

    style BOOSTED fill:#4a6fa5,color:white
    style VB fill:#7c3aed,color:white
    style AS fill:#f39c12,color:white
    style SR fill:#d63031,color:white
    style FINAL fill:#00b894,color:white

See the Salience & Persona Profiles page for full details on persona configuration.

Novelty Computation¶

How It Works¶

Novelty is computed using the nearest-neighbor distance in working memory — the minimum L2 distance between the incoming embedding and all existing working memory slots:

flowchart LR
    NEW["New memory\nembedding vector"] --> SCAN["SIMD scan of\nworking memory\n~0.5ms for 100×768"]
    SCAN --> DIST["Min L2 distance\nto nearest neighbor"]
    DIST --> NORM["Normalize to 0-1\nd / 2.0, capped"]

    style SCAN fill:#0984e3,color:white
    style NORM fill:#00b894,color:white

A high distance means the memory is genuinely novel — it's far from everything the agent has seen recently.

Normalization¶

The raw distance is normalized to [0, 1] via:

Where 2.0 is a configurable threshold representing "maximally novel."

Ingestion Hints¶

The LLM provides hints at ingestion time with three signals:

Novelty is computed automatically from working memory — the LLM does not provide it.

Safety Features¶

• Clamping: All values are clamped to [0.0, 1.0] on input
• Fallback: When no hints are provided, the system falls back to novelty-only mode (backward compatible)
• Gaming detection: If all hints are maximal (I=1.0, C=1.0, U=1.0), a warning is logged

Configuration¶

ICNU Fusion Weights¶

Custom weights can be configured via the builder:

SpectorMemory.builder()
    .icnuWeights(interest: 0.4, challenge: 0.1, novelty: 0.3, urgency: 0.2)
    .build()

Salience Profile Weights¶

Salience profiles can override ICNU weights per-user when the tenant policy allows:

{
  "icnuWeights": { "I": 0.40, "C": 0.10, "N": 0.30, "U": 0.20 },
  "interests": ["database performance", "security"],
  "persona": {
    "valenceBias": -40,
    "arousalSensitivity": 1.8
  }
}

Built-in Weight Presets¶

Weight Auto-Normalization¶

Weights are automatically normalized on construction. For example, weights of (1, 1, 1, 1) become (0.25, 0.25, 0.25, 0.25).

Worked Example¶

Agent ingests: "User has a production outage — database connections exhausted"

Stage 1: ICNU Fusion

With default weights:

Stage 2: Salience Profile Boost

The user has a salience profile with "database performance" as a CRITICAL interest:

Stage 2b: Persona Modulation

The user has a pessimistic valence bias (−40) — this memory describes a failure/outage, so the negative valence aligns with the bias, providing an additional decay resistance boost.

Final importance: 10.0 — this memory will be pinned as a flashbulb event.

MCP Integration¶

When using the MCP tools, importance fusion happens automatically if the ingestion tool provides hints:

{
  "name": "memory_remember",
  "arguments": {
    "id": "outage-2024-01",
    "text": "Production database connections exhausted at 2AM",
    "tags": "production,database,outage",
    "hints": {
      "interest": 0.7,
      "challenge": 0.5,
      "urgency": 1.0
    }
  }
}

Next Steps¶

• Salience & Persona Profiles — topic interests, persona modulation, hierarchical merge
• Dopamine — Surprise Detection — the biological novelty model
• Scoring Pipeline — the 6-phase SIMD scoring engine
• Cognitive Profiles — how profiles interact with importance