An LLM is stateless — it has no memory of its own beyond what's sitting in its context window right now. Everything an agent appears to "remember," from a name mentioned three turns ago to a preference from last week, is memory that had to be deliberately engineered around it. Agent memory is the practical skill of designing that layer: what to keep, where to keep it, and when to bring it back.
This module starts with the fundamentals: the distinction between short-term memory (the live conversation) and long-term memory (what survives across sessions), and the fuller framework of memory types — working, episodic, semantic, and procedural — that most production systems are actually built from. It then turns hands-on, building semantic memory and persistent, database-backed memory into a real LangGraph agent. It closes with a full conference talk that ties semantic, episodic, and procedural memory together into a working, memory-aware email assistant.
Watch the lessons in order, then continue on to Agent Architectures.
Googlers Aja Hammerly and Jason Davenport lay out short-term, long-term, and working memory for LLMs in plain terms, with examples of how each shows up in a real agent toolkit.
IBM's Martin Keen breaks agent memory into four distinct kinds — working, episodic, semantic, and procedural — and explains why conflating them is the most common memory-design mistake.
A practical walkthrough of LangMem's semantic memory: how to let an agent extract and store facts, relationships, and user context so it can recall them across conversations.
A hands-on build giving a LangGraph agent real persistence — wiring up a Postgres-backed checkpointer so conversation state survives restarts instead of living only in memory.
LangChain co-founder and CEO Harrison Chase ties it all together — semantic, episodic, and procedural memory, hot-path vs. background updates — while building a personalized email assistant end to end.