LangMem

LangMem is an open-source Python library from the LangChain team that provides memory primitives specifically engineered for long-horizon AI agent workflows. While LangChain and LangGraph already handle conversational state within a single session, LangMem extends that with persistent, cross-session memory that allows agents to remember facts, user preferences, prior conversations, and learned procedures across time. It addresses one of the most persistent gaps in production LLM systems: agents that lose context the moment a session ends, forcing users to re-explain themselves on every interaction.

The library offers two main usage patterns. The first is a set of functional, stateless primitives — including memory managers and prompt optimizers — that developers can integrate directly into any LangChain or LangGraph agent. These let an agent extract structured information from a conversation, decide what to write, update, or delete from long-term storage, and reflect on past interactions to improve future ones. The second pattern is a stateful storage-backed API that plugs into LangGraph's BaseStore interface, supporting in-memory, Postgres, or other persistent backends out of the box. This gives developers a clean separation between memory logic (what to remember) and storage (where to keep it).

LangMem distinguishes between several memory types modeled after cognitive science: semantic memory (facts and knowledge about the user or domain), episodic memory (specific past events and conversations), and procedural memory (learned rules, prompts, or behaviors that improve over time). The procedural-memory side is particularly novel — it includes a prompt optimizer that uses agent feedback or evaluation traces to iteratively rewrite an agent's system prompt, effectively letting the agent learn from experience without fine-tuning. Memory operations can be triggered either in the 'hot path' (synchronously during a turn) or in the 'background' (asynchronously after a turn ends), giving developers control over the latency-vs-completeness tradeoff.

LangMem is MIT-licensed and integrates tightly with the broader LangChain ecosystem, including LangGraph for orchestration and LangSmith for observability. It is best suited for teams already building on LangChain who need durable agent memory, chatbots with persistent user profiles, or copilots that improve through use. Because it is a library rather than a managed service, teams are responsible for choosing and operating their own storage backend, designing schemas, and tuning when memories are written and retrieved.