Definition
Agent memory is the set of mechanisms that let an LLM-driven agent hold onto information across turns, sessions, and tasks rather than forgetting everything once its context window fills. Because a model's context is finite and stateless between calls, durable memory must live outside the prompt and be selectively retrieved when relevant. Practitioners usually divide it into short-term and long-term memory.
Short-term versus long-term
Short-term memory is the agent's working memory: the tokens currently inside the context window, including the recent conversation and intermediate results for the task at hand. It is fast but bounded. Long-term memory persists outside the model, commonly in a vector database or structured store, and is queried so that only the few facts needed for the current decision are pulled back into the window, the same retrieval pattern used in RAG.
Categories of long-term memory
Long-term memory is frequently split into semantic memory (general facts and world knowledge), episodic memory (records of specific past interactions and their outcomes), and procedural memory (rules and learned procedures for how to act). Good memory management decides what to write, when to retrieve, and when to forget, which strongly shapes how reliably an agent behaves.
Memory is what lets an agentic workflow accumulate experience instead of starting cold each run. For a sovereign self-hoster, keeping these stores on your own hardware means an agent's accumulated knowledge about you never leaves your control. D-Central documents memory architecture as a core concern for locally run AI.
In Simple Terms
Agent memory is the set of mechanisms that let an LLM-driven agent hold onto information across turns, sessions, and tasks rather than forgetting everything once…