Glossary
agent memory
The persistent state an agent carries across turns and sessions, ranging from session-scoped scratchpads to long-term knowledge bases the agent reads and writes itself.
The state an agent reads back later. Three tiers usually distinguished:
short-term (the model’s current context window, holding the active
turn), session memory (a scratchpad across turns within one conversation
or task), and long-term memory (an external store of facts, preferences,
past outcomes the agent retrieves selectively).
Long-term memory is RAG applied inside-out. Instead of retrieving from
a static corpus, the agent retrieves from a store it has written to
itself. The hard problems are write-side: when to summarize, what to
deduplicate, how to update versus append, how to garbage-collect.
Open implementations: MemGPT (LLMs as memory-managing OS), mem0, Letta (productized MemGPT), and the memory primitives in LangChainagentsThe earliest widely-adopted LLM agent and RAG orchestration framework (2022), now with the LangGraph extension for stateful multi-step agent workflows. Open full entry and LlamaIndexretrieval-memoryAn open-source RAG framework focused on connecting LLMs to external data, with strong document-ingestion tooling and a smaller surface area than LangChain. Open full entry . None has emerged as the standard; the design space is still moving as agents move from single-turn to long-running.