# What is Working Memory?

Also called Scratchpad, Task State.

Working memory is the information an AI system actively holds while performing a task, including the current goal, intermediate results, tool outputs, and constraints being tracked. It is scoped to the task rather than to the session, and it is assembled fresh for each model call from the pieces the next step requires.

The concept describes the scratch space a multi step process needs. An agent breaking a request into steps must carry forward the plan, the results already obtained, and the conditions it must honor. Because the model is stateless, that scratch space is materialized in the prompt or in a structured state object the runtime manages.

Working memory is distinguished from conversation history by purpose. History records what was said, while working memory records what is currently true about the task: three of five records processed, this filter applied, this constraint still unmet. Keeping it structured rather than buried in prose makes it far easier to update reliably.

Capacity is the constant pressure. Long tool outputs and repeated intermediate results fill the window quickly, and quality degrades as the prompt grows. Mitigations include summarizing completed steps, storing bulky results externally and passing references to them, and pruning anything the remaining steps cannot use. Some frameworks call this context engineering.

Boundaries between working, short term, and long term memory are drawn inconsistently across frameworks and papers, and the terms are sometimes used interchangeably. The functional distinction that survives is scope: the current task, the current session, and everything beyond it. Naming should be checked against a given system's own documentation.

## Key points

- Task scoped scratch space for goals and intermediate results.
- Records what is true now, not what was said.
- Structured state updates more reliably than prose.
- Bulky results move outside and pass by reference.
- Framework definitions of this term differ.

## In practice

An agent asked to reconcile forty invoices holds a working state listing the goal, the account filter, the count processed, and three mismatches found so far. Each model call receives that state plus the next invoice, not the entire history of forty tool calls. The prompt stays small while the task runs through to completion.

## Related terms

- [Short Term Memory](/en/glossary/short-term-memory)
- [Agent Memory](/en/glossary/agent-memory)
- [Context Injection](/en/glossary/context-injection)
- [Episodic Memory](/en/glossary/episodic-memory)
- [Long Term Memory](/en/glossary/long-term-memory)

[Back to the AI Glossary](/en/glossary)