pub struct SemanticConfig {
pub enabled: bool,
pub recall_limit: usize,
pub vector_weight: f64,
pub keyword_weight: f64,
pub temporal_decay_enabled: bool,
pub temporal_decay_half_life_days: u32,
pub mmr_enabled: bool,
pub mmr_lambda: f32,
pub importance_enabled: bool,
pub importance_weight: f64,
pub embedding_provider: Option<ProviderName>,
pub embed_timeout_secs: u64,
}Expand description
Semantic (vector) memory retrieval configuration, nested under [memory.semantic] in TOML.
Controls how memories are searched and ranked, including temporal decay, MMR diversity re-ranking, and hybrid BM25+vector weighting.
§Example (TOML)
[memory.semantic]
enabled = true
recall_limit = 5
vector_weight = 0.7
keyword_weight = 0.3
mmr_lambda = 0.7Fields§
§enabled: boolEnable vector-based semantic recall. Default: true.
recall_limit: usize§vector_weight: f64§keyword_weight: f64§temporal_decay_enabled: bool§temporal_decay_half_life_days: u32§mmr_enabled: bool§mmr_lambda: f32§importance_enabled: bool§importance_weight: f64§embedding_provider: Option<ProviderName>Name of a [[llm.providers]] entry to use exclusively for embedding calls during
memory write and backfill operations. A dedicated provider prevents embed_backfill
from contending with the guardrail at the API server level (rate limits, Ollama
single-model lock). Falls back to the main agent provider when None.
embed_timeout_secs: u64Timeout in seconds applied to every embed() call inside zeph-memory.
Applies to all embedding call sites: admission control, quality gate, recall,
summarization, graph retrieval, consolidation, and tree consolidation.
Set to a higher value when using slow remote embedding providers.
Default: 5.