Struct WriteInferenceConfig 

pub struct WriteInferenceConfig {
    pub contradiction_threshold: f32,
    pub obsolete_threshold: f32,
    pub related_min: f32,
    pub related_max: f32,
    pub correction_threshold: f32,
    pub confidence_decay_factor: f32,
    pub confidence_floor: f32,
    pub value_update_enabled: bool,
    pub value_update_min_similarity: f32,
    pub value_update_prefix_share: f32,
    pub value_update_max_tail: usize,
}

Configuration for write-time inference thresholds.

Fields

contradiction_threshold: f32

Similarity above which two memories may contradict (default: 0.95).

obsolete_threshold: f32

Similarity above which an older memory is marked obsolete (default: 0.85).

related_min: f32

Minimum similarity for creating a Related edge (default: 0.6).

related_max: f32

Maximum similarity for creating a Related edge (default: 0.85).

correction_threshold: f32

Minimum similarity for a Correction to supersede (default: 0.5).

confidence_decay_factor: f32

Multiplier applied to original confidence on correction (default: 0.5).

confidence_floor: f32

Minimum confidence after decay (default: 0.1).

value_update_enabled: bool

Whether the value-update rule runs: a new memory that shares an older same-owner memory’s sentence frame (long common token prefix) but ends in a different value supersedes it, so “favorite coffee is a cortado” followed by “favorite coffee is a flat white” keeps only the correction. This is the cheap, LLM-free slice of contradiction handling; unlike bare cosine (which cannot tell paraphrase from contradiction and was removed at ~0% precision), the shared-frame test only fires on the update shape.

value_update_min_similarity: f32

Minimum embedding similarity for the value-update rule (default: 0.3).

This is a loose topical floor, not the primary gate. A value correction inherently drives cosine down (the changed value is most of a short fact’s meaning), so real corrections land well below a high cosine bar: measured on Titan V2 at 256 dims, “favorite coffee is a cortado” vs “… a flat white” is 0.61 and “phone is 1234” vs “… 4321” is 0.51; on OpenAI text-embedding-3-small they are 0.80 and 0.79. A 0.88 gate (the first default) was unreachable on either embedder, and even 0.5 sat inside the real-pair range and silently gated identical-frame pairs in production (observed on Cohere 1024: “uses Next.js for the front end” vs “uses React for the front end” never superseded). The embedder-independent is_value_update shared frame test carries the precision here; the floor only rejects genuinely off-topic pairs that a shared token frame collided on, which sit near 0.2 and below.

value_update_prefix_share: f32

Minimum fraction of the longer memory’s tokens that must be shared frame, the common prefix plus the common suffix around the changed value (default: 0.6).

value_update_max_tail: usize

Maximum differing value tokens per side (default: 4). Larger middles mean genuinely different statements, not a value change.

Trait Implementations

impl Clone for WriteInferenceConfig

fn clone(&self) -> WriteInferenceConfig

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for WriteInferenceConfig

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

Formats the value using the given formatter.

impl Default for WriteInferenceConfig

fn default() -> Self

Returns the “default value” for a type.