The Four Types of Agent Memory
Working, short-term, long-term, and org-wide memory — what each type stores, its lifetime, and implementation.
Last updated
After this section you can
- Keep the two lenses on memory apart — operational layers versus cognitive types
- Name the four layers and give each its lifetime, latency, and storage
- Recognise the tiered-memory pattern as the same idea your OS already uses
The Four Types of Agent Memory
An LLM is stateless — it forgets everything the instant a request returns. “Memory” is the engineering you bolt on to fake continuity. Two lenses on it: where memory lives (operational layers) and what it stores (cognitive types).
The model has no memory. None. Every impression of an agent that “remembers you” is your system re-assembling the right information into the context window on each turn. So agent memory is really a context-engineering problem: given a finite, expensive window, what do you load, from where, and when? Get this wrong in one of two directions and the agent either has amnesia (forgets what matters) or dementia (drowns in stale context). The four-layer model below is how you get it right — each layer tuned to a different lifetime and latency.
Two lenses — don’t confuse them
“Memory” gets used for two orthogonal ideas, and interviewers listen for whether you can separate them: