What is Long-term memory (agents)?

Long-term memory is the part of an agent's memory that outlives any single session. Where short-term or working memory lives in the context window and is wiped when the conversation ends, long-term memory is written to durable storage, a database, a vector store, a dedicated memory server, and stays available for every future task. It is the foundation of an agent that improves over time: facts, decisions, preferences, and the lessons of past attempts are recorded once and recalled when relevant, instead of being relearned from scratch each session. Long-term memory is commonly split into semantic memory (general facts and conventions) and episodic memory (specific past events), and a complete system captures both. The hard part is rarely storage, it is retrieval: because the durable store grows large while the context window stays small, the system must select only what is relevant to the current task rather than dumping everything back in. Delivered as an MCP server, long-term memory becomes a tool the model calls to record and recall context, and when the store is shared across a team, every agent benefits from what any one of them learned. This is the design Glen implements: durable, organization-scoped memory retrieved on demand through a single MCP tool.