LangMem is a toolkit that helps agents learn from their interactions over time. It extracts important information from conversations, refines agent behavior through prompt optimization, and maintains long-term memory across sessions. It provides functional primitives that work with any storage system as well as native integration with LangGraph's long-term memory store, which is available by default in LangGraph Platform deployments.
Memory can be managed in two ways. In the hot path, create_manage_memory_tool and create_search_memory_tool give the agent tools to record and search memories during a conversation, persisted through a LangGraph BaseStore such as InMemoryStore for development or AsyncPostgresStore for production. In the background, a memory manager extracts, consolidates, and updates knowledge without blocking the conversation. Guides cover semantic memory collections, user profiles, episodic memories, dynamic namespaces, message-history summarization, and use in custom agents or CrewAI.
Features
- Core memory API: storage-agnostic primitives for extracting and updating memories
- Hot-path tools: manage and search memory tools the agent calls during active conversations
- Background memory manager: automatic extraction, consolidation, and updates outside the request path
- Memory types: semantic collections, user profiles, and episodic memories
- Prompt optimization: refine single or multiple prompts from interaction history
- LangGraph store integration: works with InMemoryStore, AsyncPostgresStore, and other BaseStore backends with vector indexing
- Namespaces: dynamic namespaces scope memories per user, organization, or agent
- Short-term memory utilities: summarize message history to manage context length
