memory engine

design

no pinecone, no elasticsearch, no langchain. custom hybrid search built on SQLite.

search pipeline

1. FTS5 keyword search — full-text search with BM25 ranking
2. vector cosine similarity — embedding-based semantic search
3. weighted hybrid merge — combines both result sets with configurable weights
4. LZ4 compression — entries over 1KB are automatically compressed

embedding cache

LRU cache with configurable size. prevents redundant embedding API calls for frequently accessed entries.

[memory]
max_entries = 10000
embedding_cache_size = 1000

agent tools

the agent interacts with memory through two tools:

• memory_store — save information with automatic embedding
• memory_recall — hybrid search across all stored entries

implementation

#[async_trait]
pub trait Memory: Send + Sync {
    async fn store(&self, key: &str, value: &str) -> Result<()>;
    async fn recall(&self, query: &str, limit: usize) -> Result<Vec<MemoryEntry>>;
    async fn forget(&self, key: &str) -> Result<()>;
}

the default SQLite implementation uses parking_lot::Mutex for thread-safe access without poisoning overhead.