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.

---

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 to produce a richer importance score.

---

The Formula

\[ \text{importance} = 0.05 + \left(\sum_{i \in \{I,C,N,U\}} w_i \cdot x_i\right) \times 9.95 \]

Where:

Signal	Symbol	Range	Source
Interest	\(x_I\)	[0, 1]	LLM-provided hint
Challenge	\(x_C\)	[0, 1]	LLM-provided hint
Novelty	\(x_N\)	[0, 1]	Computed from working memory scan
Urgency	\(x_U\)	[0, 1]	LLM-provided hint

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

Weight	Default	Rationale
\(w_I\) (interest)	0.30	Agent engagement is a strong signal
\(w_C\) (challenge)	0.10	Complexity is less important than novelty
\(w_N\) (novelty)	0.40	Novelty is the strongest predictor of future usefulness
\(w_U\) (urgency)	0.20	Time-sensitive information needs priority

Output Range

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

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

---

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:

float nearestDist = workingStore.nearestDistance(quantizedVector, mins, scales);

nearestDistance() performs a SIMD-accelerated scan of all working memory slots (~0.5ms for 100 slots × 768 dims) and returns the minimum L2 distance. 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:

\[ \text{noveltyNorm} = \min\left(\frac{d_{\text{nearest}}}{2.0}, 1.0\right) \]

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

---

IngestionHints

The LLM provides hints via the IngestionHints record:

// At ingestion time
var hints = new IngestionHints(
    0.8f,   // interest: agent finds this very interesting
    0.3f,   // challenge: moderate complexity
    0.9f    // urgency: high time sensitivity
);

// Novelty is computed automatically from working memory
cognitiveTarget.ingestCognitive(id, text, type, tags, source, hints);

Safety Features

• Clamping: All values are clamped to [0.0, 1.0] on construction
• Fallback: IngestionHints.NONE triggers novelty-only mode (backward compatible)
• Gaming detection: If all hints are maximal (I=1.0, C=1.0, U=1.0), a WARN is logged

NONE Fallback

When no hints are provided (IngestionHints.NONE), the system falls back to IcnuWeights.NOVELTY_ONLY — importance is determined solely by nearest-neighbor distance, matching the pre-ICNU behavior.

---

Configuration

Fusion Weights

var memory = SpectorMemory.builder()
    .icnuWeights(new IcnuWeights(0.4f, 0.1f, 0.3f, 0.2f))  // custom weights
    .build();

Built-in Weight Presets

Preset	I	C	N	U	Use Case
DEFAULT	0.30	0.10	0.40	0.20	General-purpose
NOVELTY_ONLY	0.00	0.00	1.00	0.00	Backward-compatible

Weight Auto-Normalization

Weights are automatically normalized on construction:

var w = new IcnuWeights(1f, 1f, 1f, 1f);
// → interest=0.25, challenge=0.25, novelty=0.25, urgency=0.25

---

Worked Example

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

Signal	Value	Source
Interest	0.7	LLM hint — agent finds this relevant
Challenge	0.5	LLM hint — moderate complexity
Novelty	0.9	Working memory scan — nothing like this recently
Urgency	1.0	LLM hint — production outage

With default weights:

\[ \text{weighted} = 0.30 \times 0.7 + 0.10 \times 0.5 + 0.40 \times 0.9 + 0.20 \times 1.0 = 0.81 \]

\[ \text{importance} = 0.05 + 0.81 \times 9.95 = \mathbf{8.11} \]

This is a high-importance memory (8.11 / 10.0) — it will be prioritized in future recalls and resist time decay.

---

MCP Integration

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

{
  "name": "core_memory_append",
  "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
    }
  }
}

Backward Compatibility

The hints field is optional. When omitted, importance is computed using novelty-only mode — identical to the pre-ICNU behavior.