Scoring Regimes & Fusion Modes¶

Spector Memory evaluates candidate memories through two primary scoring regimes: the Generic Scoring Regime and the Soul-Conditioned Scoring Regime (FERS).

1. Generic Scoring Regime¶

The Generic regime operates on standard vector similarity and metadata features when no SoulContext is provided or when AISME is disabled.

1.1 Score Fusion Modes¶

Spector supports two fusion modes configured via RecallOptions (ScoreFusionMode) or spector.memory.recall.score-fusion-mode:

MULTIPLICATIVE (Default)¶

Preserves backwards-compatible composite scoring where semantic tag overlap acts as a multiplicative boost:

Where \(\text{BaseScore} = \text{Similarity} \cdot \text{ImpDecayFactor} \cdot \text{ValenceMultiplier}\).

ADDITIVE¶

Applies linear convex combination governed by the live parameter \(\alpha \in [0.0, 1.0]\):

1.2 Early Graph Associative Prior (Phase 6 Fusion)¶

When enabled (enableAssociativePrior = true), an \(O(1)\) Hebbian and co-activation associative prior \(A_g \in [0.0, 1.0]\) modulates candidates in Phase 6 fusion:

• Multiplicative: \(\text{Score} = \text{Score}_{\text{base}} \cdot (1.0 + \delta \cdot A_g)\)
• Additive: \(\text{Score} = \text{Score}_{\text{base}} + \delta \cdot A_g\)

2. Soul-Conditioned Regime (FERS)¶

When a SoulContext or AISME bundle is present on the RecallSignal, Spector activates the Free-Energy Resonance Scorer (FERS).

2.1 Formula¶

Where: - \(\text{Sim}\) is the INT8 / SVASQ vector similarity. - \(\sigma(\Delta F) = \frac{1}{1 + e^{-\Delta F}}\) is the logistic sigmoid squashing variational free energy delta. - \(\text{Resonance}\) is affective and goal resonance with the persona's active mental state.

2.2 Dynamic Weight Modulation & Hysteresis Damping¶

\(\alpha, \beta, \gamma\) are dynamically computed per query by SoulConditionedWeightProvider from: 1. Goal Relevance — active task alignment. 2. Value / Purpose Alignment — core persona drives. 3. Cognitive Profile — modulates precision (e.g. HYPERFOCUS, DIVERGENT).

To prevent perceptual oscillation between consecutive turns, weights undergo Exponential Moving Average (EMA) hysteresis damping with rate clamping:

Where \(\lambda = 0.15\) (decay factor), with deadband \(\epsilon = 0.01\) and max slew rate per cycle.

3. Lateral Inhibition & Interference Resolution¶

In Phase 8 of recall, top candidates are evaluated for mutual interference by LateralInhibitionRelay:

1. Clustering: Single-linkage clustering over cosine similarity \(\theta \ge 0.88\).
2. Soft Redundancy Attenuation: Non-top candidates in redundant clusters receive a graded penalty \(\text{penalty}_i = 1.0 - \kappa_{\text{soft}} \cdot (1.0 - 1.0 / \text{rank}_i)\) without being dropped.
3. Hard Contradiction Arbitration: Candidates flagged with FLAG_CONTRADICTED are scored against confidence \(C_i = 0.4 \cdot \text{recency} + 0.3 \cdot \text{corroboration} + 0.3 \cdot \text{storageStrength}\). Losers receive \(\kappa_{\text{hard}} = 0.30\).

4. Telemetry & Explainability¶

Every recall returns detailed scoring provenance in ScoreBreakdown: