LangMem
Type: full-code · Vendor: LangChain · Language: Python · License: MIT · Status: active · Status in practice: emerging
LangChain's long-term-memory SDK that lets agents store, search, and update semantic, episodic, and procedural memories outside the prompt window.
Description. LangMem is a Python library from the LangChain team that adds long-term memory primitives to agent applications. Its conceptual guide names three memory types — semantic (facts that ground responses), episodic (successful past interactions used as learning examples), and procedural (how the agent should behave and respond) — and exposes typed extraction, storage, and prompt-optimization APIs over each. LangMem integrates with LangGraph's BaseStore so writes can be managed automatically and surfaced as tools the agent calls directly, rather than only as a sidecar service.
Agent loop shape. Library invoked from inside an agent's tool-calling loop, optionally backed by LangGraph's BaseStore. The agent (or a background manager) calls LangMem extractors that classify writes into semantic, episodic, and procedural buckets; semantic and episodic memories are persisted as searchable items, while procedural memory is materialised as updated system-prompt instructions via a prompt-optimizer (metaprompt / gradient / prompt_memory). Retrieved memories are injected back into the next prompt.
Primary use cases
- persistent agent memory across LangGraph runs
- episodic recall of prior successful interactions
- semantic fact stores that agents update over time
- procedural memory for evolving agent system prompts
Key concepts
- Semantic memory → semantic-memory (docs) — Facts and relationships that ground the agent's responses; extracted from conversations.
- Episodic memory → episodic-memory (docs) — Successful past interactions kept as learning examples that guide future behavior.
- Procedural memory → procedural-memory (docs) — How the agent should behave and respond; materialised by updating the system prompt via an optimizer.
- Prompt optimizer (metaprompt / gradient / prompt_memory) (docs) — Three optimizer strategies that update system prompts from past conversations and feedback.
- LangGraph BaseStore integration (docs) — Backing store used to persist extracted memories automatically across agent runs.
Patterns this full-code implements —
- ★★Episodic Memory
LangMem's conceptual guide defines episodic memory as preserving successful interactions as learning examples, and there is a dedicated extract_episodic_memories guide.
- ★Semantic Memory
LangMem's conceptual guide names semantic memory as one of three memory types and ships an extract_semantic_memories guide for the collection pattern.
- ★Procedural Memory
Procedural memory is the third memory type in LangMem's taxonomy and is materialised by the prompt-optimizer, which updates system prompts from conversation logs and feedback.
References
“Semantic memory stores the essential facts and other information that ground an agent's responses.”
“Update system prompts based on conversation logs or agent trajectories. Feed past conversations and feedback into an optimizer to fix common issues like missed requirements or poor response structure.”
Provenance
- Last analyzed:
- Last updated:
- Verification status: verified