An Overview of Modern Memory Management Architectures in LLM Agents (vinithavn.medium.com)

A concise primer maps modern memory architectures that make LLM agents persistent, personalized, and scalable beyond fixed context windows. The article breaks memory into short-term (working/context window), procedural (system prompts/rules), and long-term (episodic and semantic) stores, and surveys practical approaches: short-term fixes like recent-message buffers and summarization; Retrieval-Augmented Generation (RAG) where queries are embedded, matched via semantic search to a vector store, and merged with active context; and RAG+Knowledge-Graph hybrids that add structured nodes/edges for richer relational retrieval. These patterns enable continuity across sessions, preference retention, and reduced hallucination by grounding responses in external memory. It then covers recent, more autonomous systems for long-term management. MemGPT proposes an OS-inspired hierarchical model with Main (fast, limited) and External (large, archival) memories and lets the LLM invoke function-like memory ops (search, swap, update) to page relevant items into context—effectively extending usable context and enabling dynamic memory editing. Langmem (in the LangChain ecosystem) provides a core memory API plus tools to create/manage/search persistent memories across backends, simplifying integration. Technical implications: better personalization and longer-horizon reasoning, but added complexity in storage, embedding quality, retrieval latency, consistency/conflict resolution, and update strategies.