Memory

By default, an agent forgets everything the moment a conversation ends. The Memory server fixes that for a narrow, useful slice of information: it stores entities, their observations, and the relations between them as a simple local knowledge graph, persisted to a JSONL file on disk.

Across sessions, an agent can create entities for people, projects, or preferences it’s learned about, attach observations to them over time, and query the graph back out when relevant context resurfaces later. Because storage is just a JSONL file, it’s easy to inspect, back up, version, or reset — set MEMORY_FILE_PATH to point it somewhere durable if you don’t want memory tied to the server’s install location.

This is a good fit for personal-assistant-style agents, long-running project agents, or anywhere you want continuity without standing up a real database. It is not a substitute for a vector store or full RAG pipeline — the graph is small, structured, and meant for facts an agent should just remember, not for retrieving over a large document corpus.