LLMs are stateless — they don't remember previous interactions. Memory systems give your AI applications the ability to maintain context across conversations and sessions.

Types of Memory

1. Buffer memory: Store the full conversation history. Simple but hits context window limits quickly.
2. Window memory: Keep only the last N messages. Trades recall for bounded context usage.
3. Summary memory: An LLM periodically summarizes the conversation, compressing old messages. Good balance of recall and efficiency.
4. Entity memory: Extract and track entities (people, places, facts) mentioned in conversation. Structured and queryable.
5. Vector store memory: Embed and store all messages, retrieve relevant ones via similarity search. Scales to unlimited history.

LangGraph State Management

LangGraph uses explicit state schemas:

from typing import Annotated, TypedDict
from langgraph.graph import add_messages

class AgentState(TypedDict):
    messages: Annotated[list, add_messages]
    user_preferences: dict
    current_task: str | None
    completed_steps: list[str]

State is passed between nodes and persisted across sessions using checkpointers:

from langgraph.checkpoint.memory import MemorySaver

checkpointer = MemorySaver()
app = graph.compile(checkpointer=checkpointer)

# Resume conversation with thread_id
config = {"configurable": {"thread_id": "user-123"}}
result = app.invoke({"messages": [new_message]}, config)

Long-Term Memory Patterns

For applications that need to remember users across weeks or months:

• User profile store: Persist learned preferences, facts, and interaction history in a database
• Episodic memory: Store summaries of past conversations, retrievable by topic
• Semantic memory: Maintain a knowledge graph of entities and relationships

Practical Tips

• Start with simple buffer memory and add complexity only when needed
• Always set a maximum context budget and handle overflow gracefully
• Test memory retrieval with realistic conversation lengths (50+ turns)
• Consider privacy — users should be able to view and delete their stored memory