Curiosity · Concept

Agentic Memory

LLMs are stateless by default: every call starts from zero context. Production agents need memory that outlives a single request. The usual stack has three layers: a short-term scratchpad held in the current context, episodic memory of past interactions (often stored in a vector database and retrieved by similarity), and semantic memory that distills recurring facts into structured records. Systems like MemGPT, LangGraph checkpoints, and Letta implement this pattern, and research like Generative Agents (Park et al. 2023) showed how reflection plus retrieval produces surprisingly rich agent behavior.

Quick reference

Proficiency
Intermediate
Also known as
agent memory, long-term agent memory
Prerequisites
agents, vector database, context window

Frequently asked questions

What is agentic memory?

Agentic memory is any mechanism that gives an LLM agent persistent state beyond the current context window — short-term scratchpads, long-term semantic stores, and episodic logs of past interactions — so the agent can learn and carry knowledge across sessions.

What are the common memory tiers?

Short-term (working memory in the active context), episodic (logs of past events retrieved by similarity), and semantic (distilled facts, preferences, entity profiles). Many systems also add procedural memory for learned skills or tool-use patterns.

How do you avoid context bloat?

Retrieve only relevant memories (vector similarity or keyword filters), summarize old conversations into compressed records, and evict or archive stale memory. MemGPT treats this as OS-style paging between context and an external store.

When do I actually need a memory system?

Single-turn or single-session workflows don't need one. Multi-session assistants, user-personalized agents, long-running research agents, and anything that should 'remember who I am' do. Start with a simple user-profile JSON and add episodic retrieval only when you hit its limits.

Sources

  1. Packer et al. — MemGPT: Towards LLMs as Operating Systems — accessed 2026-04-20
  2. Park et al. — Generative Agents: Interactive Simulacra of Human Behavior — accessed 2026-04-20

Related

  • Context Window — The context window is the maximum number of tokens — prompt plus output — a language model can process in a single call, bounded by architecture and memory.
  • Reflexion (Self-Reflection Loop) — Reflexion is an agent pattern where, after an attempt fails, the LLM writes a natural-language self-critique of what went wrong and stores it in episodic memory so the next attempt is better informed — learning by reflection instead of gradient descent.
  • Retrieval-Augmented Generation (RAG) — Retrieval-Augmented Generation (RAG) is a pattern where an LLM is grounded on retrieved passages at query time — fewer hallucinations, up-to-date answers, no retraining required.
  • All Concepts — Browse the complete concepts library
  • VSET — Programmes offered — B.Tech programs (CSE, AI&ML, AI&DS, Electronics, IIoT)