Skip to main content

Memory System

A Letta-inspired dual-tier memory architecture enabling AI agents to maintain working memory and unlimited long-term storage with semantic recall.

Overview​

Core Memory: Always-visible working memory (4 blocks: persona, user, facts, context) Archival Memory: Unlimited searchable storage with vector embeddings Automatic Management: System extracts facts without explicit agent commands

Built on LanceDB for efficient vector storage and search.

Quick Example​

from suzent.memory import MemoryManager, LanceDBMemoryStore

# Initialize
store = LanceDBMemoryStore(uri=".suzent/data/memory")
await store.connect()

manager = MemoryManager(
    store=store,
    embedding_model="text-embedding-3-large",
    llm_for_extraction="gpt-4o-mini"
)

# Use
blocks = await manager.get_core_memory(user_id="user-123")
results = await manager.search_memories("user preferences", user_id="user-123")

Key Features​

  • ✅ Core memory blocks (persona, user, facts, context)
  • ✅ Semantic + full-text hybrid search
  • ✅ Automatic LLM-based fact extraction
  • ✅ Rich Context Awareness (captures extracted intent & outcomes)
  • ✅ Auto-summarizing Core Memory (keeps "facts" block fresh)
  • ✅ Importance scoring and deduplication
  • ✅ Thread-safe agent tools