Agent Memory

Agent Memory refers to the persistent storage and retrieval infrastructure that allows an LLM-based AI agent to maintain, recall, and reason over information across sessions, conversations, and task executions. Unlike the bounded context window of an LLM (which limits how much information the model can directly attend to within a single inference), agent memory operates as an external store that the agent can write to, read from, and search over time. Agent memory systems address three core challenges: (1) unified representation — supporting diverse memory types (conversations, entities, events, facts, preferences) in a single framework; (2) efficient hybrid retrieval — combining vector similarity, keyword matching, and graph traversal to find relevant memories without cross-database orchestration; and (3) accurate denoised retrieval — filtering, deduplicating, and resolving conflicts in retrieved memories before they enter the LLM context.

The mechanics typically involve a write path that converts agent experiences into structured memory items (often using an LLM to extract facts, entities, and relationships), a storage layer that organizes those items into a queryable format (vector indexes, knowledge graphs, or hybrid structures), and a read path that retrieves relevant items for each new query—often with denoising, conflict resolution, and token-budget control. Modern systems include Mem0 (Apr 2025), MemOS, Memoria, Mandol (Microsoft/ISCAS, Jun 2026), Zep, and a long lineage of retrieval-augmented memory architectures.

Agent memory is distinct from long-term-memory (the broader concept from cognitive science) and from rag (the foundational retrieval-augmented-generation framework). It is the application of RAG-like retrieval to an agent's own accumulated experiences rather than to a static corpus. Closely related to parametric-memory, episodic-memory, agent-harness, and world-model.