Agent Memory

System agents published

Also known as: Memory System, Agent State Persistence

Definition

Mechanisms that allow an agent to retain and access information beyond the current context window. Memory enables agents to learn from past interactions, maintain state across sessions, and reference earlier observations. Memory systems range from simple conversation history to sophisticated retrieval- augmented stores.

What this is NOT

  • Not the same as context window (memory is external storage, context is active prompt)
  • Not just chat history (memory systems can be selective, summarized, or structured)
  • Not the model's weights (those are frozen after training)

Alternative Interpretations

Different communities use this term differently:

llm-practitioners

External storage that an agent can read from and write to, enabling persistence beyond the LLM's context window. Implementations include vector databases, key-value stores, and structured databases accessed via retrieval.

Sources: LangChain Memory documentation, MemGPT paper (2023), Letta documentation

cognitive-science

The cognitive systems that encode, store, and retrieve information. Distinctions include: working memory (active, limited), episodic memory (events), semantic memory (facts), procedural memory (skills).

Sources: Cognitive psychology literature, Baddeley's working memory model

Examples

  • MemGPT managing a virtual context with memory pagination
  • A customer service agent remembering past support tickets for a user
  • A coding agent that remembers project structure and past decisions
  • A personal assistant that builds up knowledge about user preferences

Counterexamples

Things that might seem like Agent Memory but are not:

  • A stateless API that treats each request independently
  • Including the full conversation in each prompt (that's context, not memory)
  • The model's training data (that's parametric knowledge, not memory)

Relations

See Also

Implementations

Tools and frameworks that implement this concept: