Struct CognitiveLlmService 

pub struct CognitiveLlmService<J: LlmJudge> { /* private fields */ }

The unified LLM intelligence layer for memory operations.

Wraps any LlmJudge implementation and provides typed judgment methods with specialized prompts for each cognitive task.

Implementations

impl<J: LlmJudge> CognitiveLlmService<J>

pub fn new(judge: J) -> Self

Create a new service backed by the given LLM judge.

pub async fn judge_invalidation( &self, old: &MemorySummary, new: &MemorySummary, ) -> Result<InvalidationVerdict, LlmJudgeError>

Determine whether a new memory invalidates an existing one.

This is the key method that replaces the cosine similarity heuristic for temporal invalidation. The LLM can understand that “Alice works at Acme” is invalidated by “Alice joined Google” even though the embeddings are quite different.

pub async fn detect_contradiction( &self, a: &MemorySummary, b: &MemorySummary, ) -> Result<ContradictionVerdict, LlmJudgeError>

Determine whether two memories contradict each other.

pub async fn resolve_entities( &self, candidates: &[EntityCandidate], ) -> Result<Vec<EntityMergeGroup>, LlmJudgeError>

Resolve a batch of entity references into merge groups.

pub async fn consolidate( &self, cluster: &[ClusterMember], ) -> Result<ConsolidationDecision, LlmJudgeError>

Decide how to consolidate a cluster of similar memories.

pub async fn canonicalize_topic( &self, message: &str, existing_topics: &[String], ) -> Result<TopicLabel, LlmJudgeError>

Map a raw user message to a canonical topic label.