LangMem

LangMem is an open-source Python SDK developed by LangChain to provide AI agents and LangGraph workflows with long-term memory. It allows agents to learn, adapt, and maintain behavioral consistency across sessions by automatically extracting, storing, and consolidating semantic facts, episodic interaction summaries, and procedural execution patterns.

graph TD
    Interaction[User & Agent Interaction] --> Collector[Conversation Collector]
    Collector --> Extractor[Background LLM Extractor & Consolidator]
    Extractor --> SemanticMem[Semantic Memory<br/>Facts & User Profile]
    Extractor --> EpisodicMem[Episodic Memory<br/>Past Task Summaries]
    Extractor --> ProceduralMem[Procedural Memory<br/>Rules & System Directives]
    SemanticMem & EpisodicMem & ProceduralMem --> BaseStore[(LangGraph BaseStore / Persistent DB)]
    BaseStore --> Tools[Hot-Path Recall & Manage Tools]
    Tools --> Agent[LangGraph Agent Graph]

Cognitive Memory Primitives

LangMem structures agent memory into three primary operational categories:

1. Semantic Memory (User & World Knowledge):
   • Maintains continuous profiles of users, preferences, and factual constraints.
   • Updates dynamically as conversations progress, combining new information with prior records.
2. Episodic Memory (Past Experiences):
   • Summarizes and indexes completed threads, debugging sessions, and tasks.
   • Allows agents to recall how past problems were solved or past decisions were justified.
3. Procedural Memory (Behavioral & Prompt Adaptation):
   • Automatically refines system prompts, instructions, and behavioral guidelines based on feedback or error corrections.

Key Characteristics

• LangGraph Native: Built to leverage LangGraph's persistent state and storage primitives (BaseStore), running seamlessly with PostgreSQL, SQLite, or in-memory stores.
• Background & Hot-Path Modes: Supports both real-time "hot path" memory tool invocations by the agent and asynchronous background memory extraction routines that run after a conversation completes.
• Pluggable & Model Agnostic: Works with any LLM provider (OpenAI, Anthropic, Gemini, Ollama) and custom embedding models.
• Open Source: Licensed permissively under MIT and fully runnable locally without external SaaS dependencies.

Python Quickstart

Install LangMem:

pip install -U langmem

Example usage extracting and retrieving user facts:

from langmem import create_memory_store, create_memory_manager

# 1. Create a memory store instance
store = create_memory_store()

# 2. Create memory manager configured for semantic extraction
manager = create_memory_manager(
    "anthropic:claude-3-5-sonnet-latest",
    store=store,
    namespace=("users", "user_42", "memories")
)

# 3. Process conversation turns and extract facts
conversation = [
    {"role": "user", "content": "I live in Berlin and always test using pytest."},
    {"role": "assistant", "content": "Understood! I'll use pytest for your tests."}
]
manager.invoke(conversation)

# 4. Search extracted memories
memories = store.search(("users", "user_42", "memories"), query="testing framework")
for mem in memories:
    print(mem.value)

Related Concepts

• LLM and Agent Memory - Foundational agent memory concepts and cognitive taxonomy.
• Mem0 - Vector-first persistent memory layer.
• Letta - Stateful agent framework with hierarchical memory.
• Graphiti and Zep - Temporal knowledge graph engine for agent memory.