Memory Store and Knowledge Management

Memory Store and Knowledge Management

Build persistent, searchable memory systems for LangGraph agents using the Store interface.

Overview

LangGraph's Store interface enables agents to:

• Remember across conversations - Persist data beyond single threads
• Search semantically - Find relevant memories using natural language
• Isolate user data - Organize memories with namespaces
• Build knowledge - Accumulate facts and concepts over time

Unlike checkpointing (which saves workflow state within a thread), the Store provides cross-thread persistence for long-term knowledge retention.

---

Core Concepts

1. Store Interface

The Store is LangGraph's abstraction for persistent memory:

from langgraph.store.memory import InMemoryStore

# Basic store (key-value only)
store = InMemoryStore()

# Store with semantic search
from langchain.embeddings import init_embeddings

embeddings = init_embeddings("openai:text-embedding-3-small")
store = InMemoryStore(
    index={
        "embed": embeddings,
        "dims": 1536,
    }
)

When to use:

• User profiles and preferences
• Facts and knowledge bases
• Historical context across sessions
• Shared data between threads

2. Namespace Organization

Namespaces provide hierarchical isolation using tuples:

# Namespace structure: (category, user_id, [subcategory])
user_facts = ("memories", "user_123")
user_prefs = ("preferences", "user_123")
global_kb = ("knowledge_base",)

# Store data
store.put(user_facts, "fact_1", {"text": "User loves Python"})
store.put(user_prefs, "pref_1", {"theme": "dark"})
store.put(global_kb, "concept_1", {"topic": "LangGraph basics"})

Namespace patterns:

• ("memories", user_id) - Per-user facts
• ("preferences", user_id) - User settings
• ("knowledge_base",) - Global shared knowledge
• ("conversations", user_id, thread_id) - Thread-specific episodic memory

3. Storing Data

Use store.put(namespace, key, value):

# Store with auto-generated key
store.put(
    ("memories", "user_123"),
    "mem_001",
    {
        "text": "User mentioned they're learning DSPy",
        "timestamp": "2026-01-13T10:00:00Z",
        "context": "onboarding conversation"
    }
)

# Store with indexing control
store.put(
    ("memories", "user_123"),
    "mem_002",
    {"text": "User completed tutorial"},
    index=False  # Don't embed this item
)