Struct ScoreWeights 

pub struct ScoreWeights {
    pub alpha: f32,
    pub beta: f32,
    pub gamma: f32,
    pub delta: f32,
    pub half_life_secs: f32,
}

Weights for the multi-signal recall score (scoring v2, shipped in v0.7.0).

score = similarity
      + max(similarity, 0)
        · ( β · recency
          + γ · (importance − 0.5)
          + δ · tanh(access_count / 5) )

recency = exp( -ln(2) · age_secs / half_life_secs )
age_secs = now - memory.created_at

Design notes (from the 2026-11-06 scoring-v2 bake-off; see the ranking_v2 harness at mnemo/examples/ranking_v2.rs):

• Similarity-gated. The recency / importance / usage terms are multiplied by max(similarity, 0), so a stale-but-semantically-off memory can’t ride recency or importance past a semantically-on-point one. Pre-v0.7.0 scoring was purely additive.
• Importance is centered at 0.5. A default-importance memory contributes zero importance bonus; importance only pushes a score up or down from that reference point.
• Usage is saturating. tanh(uses/5) caps the feedback bonus so a heavily-returned memory can’t keep accumulating weight past a plateau. At 5 uses the bonus is ~76% of its asymptote; at 10, ~96%.
• Recency is decoupled from the feedback loop. Age is measured from created_at, not accessed_at, so a recall() call does not bump its own results’ recency scores on later queries.

The output of this scorer is on a different scale from the pre-v0.7.0 formula. Callers that compared raw scores to a fixed threshold should re-pick the threshold against the new scale.

Fields

alpha: f32

Weight on semantic similarity. Kept in the struct for calibration hooks and symmetry with the other weights; scoring v2 always uses it as 1.0 because the leading similarity term is ungated and the whole score sits on the cosine-similarity scale by design.

beta: f32

Weight on the recency bonus (multiplied by max(sim, 0)).

gamma: f32

Weight on the importance bonus (importance − 0.5), centered at the default-importance 0.5 (multiplied by max(sim, 0)).

delta: f32

Weight on the saturating usage bonus tanh(access_count / 5) (multiplied by max(sim, 0)).

half_life_secs: f32

Half-life of the recency term, in seconds.

Implementations

impl ScoreWeights

pub fn score( &self, similarity: f32, age_secs: f32, importance: f32, access_count: u32, ) -> f32

Combine the four signals into a single score. See the struct docs for the full formula and the scoring-v2 rationale.

Trait Implementations

impl Clone for ScoreWeights

fn clone(&self) -> Self

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Copy for ScoreWeights

impl Debug for ScoreWeights

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Default for ScoreWeights

fn default() -> Self

Returns the “default value” for a type.