Memory and Knowledge

Agent memory systems divide into distinct types that mirror cognitive science categories, each solving a different persistence and retrieval challenge: short-term memory holds conversation history within a session, working memory tracks in-progress task state through scratchpads and variables, long-term memory persists facts across sessions in vector databases or knowledge bases, and episodic memory logs past interactions so agents can learn from previous experience. Understanding these categories is foundational because choosing the right memory type for each piece of information determines whether your agent maintains coherent state across complex multi-step tasks.

Episodic memory stores specific past experiences tied to a time and context (what happened during a particular debugging session, how the agent resolved a specific error), while semantic memory stores general, abstracted knowledge (project conventions, API patterns, architectural decisions) that the agent applies across many future tasks. Episodic memories are most useful for avoiding repeated mistakes: an agent that recalls "last time I tried this approach it failed and required a full rollback" can make better decisions than one that starts fresh every session.

Short-term memory is the conversation history an agent carries within a single session: the messages array passed to each API call, which persists across turns but resets when the session ends. Without it, every agent response would be stateless and context-blind, making multi-step tasks impossible since the agent would forget what it was doing between each tool call.