Struct AdmissionControl 

pub struct AdmissionControl { /* private fields */ }

A-MAC adaptive memory admission controller (#2317).

Evaluates five factors (future utility, factual confidence, semantic novelty, temporal recency, content-type prior) and rejects messages below the configured composite score threshold before they are persisted.

Optionally extended with a goal-conditioned write gate (#2408) that adds a sixth factor based on the cosine similarity between the current goal embedding and the candidate memory.

Examples

use zeph_memory::{AdmissionControl, AdmissionWeights};

let weights = AdmissionWeights {
    future_utility: 0.3,
    factual_confidence: 0.2,
    semantic_novelty: 0.2,
    temporal_recency: 0.1,
    content_type_prior: 0.2,
    goal_utility: 0.0,
};
let controller = AdmissionControl::new(0.4, 0.1, weights);

Implementations

impl AdmissionControl

pub fn new( threshold: f32, fast_path_margin: f32, weights: AdmissionWeights, ) -> Self

Create a new admission controller.

• threshold — composite score [0, 1] below which messages are rejected.
• fast_path_margin — when all non-LLM factors already push the score far above the threshold (by at least this margin), the LLM future_utility call is skipped.
• weights — factor weights; normalized automatically so they sum to 1.0.

pub fn with_provider(self, provider: AnyProvider) -> Self

Attach a dedicated LLM provider for future_utility evaluation.

When set, this provider is used instead of the caller-supplied fallback.

pub fn with_goal_gate(self, config: GoalGateConfig) -> Self

Enable goal-conditioned write gate (#2408).

pub fn threshold(&self) -> f32

Return the configured admission threshold.

pub async fn evaluate( &self, content: &str, role: &str, fallback_provider: &AnyProvider, qdrant: Option<&Arc<EmbeddingStore>>, goal_text: Option<&str>, ) -> AdmissionDecision

Evaluate admission for a message.

goal_text: optional current-turn goal context for goal-conditioned scoring. Ignored when the goal gate is disabled or goal_text is None/trivial (< 10 chars).

Fast path: skips LLM when heuristic-only score is already above threshold + fast_path_margin. Slow path: calls LLM for future_utility when borderline.

On LLM failure, future_utility defaults to 0.5 (neutral).