Agent Memory (Short-term / Long-term)

Agent memory is typically divided into two categories that mirror human cognition. **Short-term memory** is the agent's active context window — the accumulating record of the current conversation, tool calls, and observations held in the LLM's prompt. It is fast, immediately accessible, and bounded: as a task grows, the context window fills, and older information must be summarized, discarded, or offloaded.

**Long-term memory** persists beyond the context window and across sessions. It is implemented through external storage systems: vector databases for semantic retrieval (the agent can query "what do I know about this customer?"), key-value stores for structured facts (user preferences, entity states), and relational databases for structured histories. When the agent needs information, it queries long-term memory and injects the relevant results into its short-term context. A third category, **episodic memory**, stores compressed summaries of past task runs, enabling agents to learn from prior successes and failures.

For the enterprise, memory architecture is a critical design decision. A customer service agent that remembers a customer's past interactions and product configuration is exponentially more valuable than one that asks the same questions every session. A financial research agent that builds a persistent knowledge graph of company relationships over months of work creates compounding value. The governance concern is privacy: long-term memory stores contain rich behavioral data that must be protected, segmented by user, and subject to deletion on request.