Struct FactStore 

pub struct FactStore { /* private fields */ }

Manages facts for one agent across all actors.

Implementations

impl FactStore

pub fn new( storage: Arc<dyn AgentStorage>, agent_id: String, config: FactsConfig, ) -> FactStore

pub async fn add_facts( &self, actor_id: &str, facts: Vec<KeyFact>, ) -> Result<Vec<KeyFact>, AgentError>

Add new facts for an actor, enforcing max_facts by evicting lowest priority rows. Returns the authoritative post-write set of facts for the actor so callers can refresh their cache without re-reading from storage.

pub async fn get_facts( &self, actor_id: &str, ) -> Result<Vec<KeyFact>, AgentError>

Load all facts for an actor.

pub async fn get_facts_by_category( &self, actor_id: &str, category: &FactCategory, ) -> Result<Vec<KeyFact>, AgentError>

Load facts filtered by category.

pub async fn query( &self, filter: &FactFilter, ) -> Result<Vec<KeyFact>, AgentError>

Query facts across actors.

pub async fn delete_fact( &self, actor_id: &str, fact_id: &str, ) -> Result<(), AgentError>

Delete a specific fact.

pub async fn delete_actor_data(&self, actor_id: &str) -> Result<(), AgentError>

Delete all data for an actor (facts + sessions). Privacy compliance.

pub fn format_for_context(&self, facts: &[KeyFact], max_tokens: usize) -> String

Format facts as a string for chat-time system prompt injection.

Output is a plain bulleted list of fact statements, ordered by priority, truncated to fit max_tokens. Category labels and confidence scores are intentionally omitted because they look like metadata noise to the chat LLM and have caused the model to hedge (“I don’t know your identity”) even when the facts clearly state the user’s name.

Use the extractor’s separate prompt format when feeding facts back into the fact extractor for dedup guidance.

pub fn agent_id(&self) -> &str

Get the agent ID this store manages.

pub fn config(&self) -> &FactsConfig

Get the underlying config.