Agent Memory

The mechanism by which an agent retains information across turns, sessions, or long-running tasks beyond the context window.

LLMs are stateless: each call only knows what’s in its context window. Agent memory is the workaround — a persistence layer that stores observations, decisions, and outcomes between calls, retrieving relevant fragments when needed.

Common forms: episodic memory (specific events: “yesterday the user asked about pricing”), semantic memory (general facts: “the user is in healthcare”), procedural memory (how-to: “for refunds, contact Stripe first”). Implementations range from a simple JSON log appended to each prompt, to vector databases retrieved via embedding similarity, to graph databases for relational facts. The right choice depends on access patterns: bounded chat history wants a log; cross-session learning wants a vector store.