WTF is Agent Memory?
Ways for an agent to remember information beyond a single context window or conversation.
Also called: long-term memory, memory
TL;DR
A model forgets everything once the conversation ends or the context fills up. Agent memory stores useful facts (user preferences, past decisions, project notes) outside the model and brings them back when relevant, so the agent improves over time instead of starting from zero.
How it works
Common split: short-term memory (the current context window) and long-term memory (external storage). Long-term memory may be a vector store queried by similarity, a structured database, or plain files the agent reads and edits. Research systems like Generative Agents (memory streams with reflection) and MemGPT (OS-style paging between context and storage) shaped the design space. Hard parts: deciding what to write, what to retrieve, and how to forget or correct stale memories.
Origin
- Coined / introduced by
- Research community
- When
- Apr 2023
- Story
- Generative Agents (Apr 2023) and MemGPT (Oct 2023) were influential early designs.