Working Memory
The small, fast scratch space an agent uses to hold what it’s thinking about right now.
Think of it like
The sticky notes on your desk during a task — grabbed, glanced at, and tossed when you’re done.
Example
While booking a trip, the agent keeps the current dates, budget, and shortlisted flights in working memory, not in long-term storage.
How it actually works
For an LLM agent, working memory usually is the context window — the live set of tokens it’s reasoning over. It’s fast but tiny and volatile: once it fills or the task ends, that content is gone unless it’s written somewhere durable. The whole discipline of context management exists because working memory is scarce and precious.
For product teams
Working memory is limited, so what an agent can juggle at once is a real constraint on task complexity.
For engineers
The active context the model reasons over each step — fast, capacity-bounded, and non-persistent.
Related
- Episodic Memory — The durable counterpart for lasting facts.
- Context window — Where working memory physically lives for an LLM.
- Scratchpad — A place to jot intermediate thinking.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome