Side by side
Context Window & Memory
The context window limits the current interaction; application memory retains information that may be selected for later interactions.
Some frameworks call conversation state short-term memory. Always distinguish retained state from the actual information supplied to the model.
Swap columns ⇄| Dimension | Context Window | Memory |
|---|---|---|
| Purpose | Bound the information available during a model interaction. | Retain information for reuse. [Anthropic][LangChain] |
| Role in the system | A model or service capacity constraint. | Application state and recall mechanisms. [Anthropic][LangChain] |
| Inputs and outputs | Supplied tokens and generated content consume available capacity. | Selected information is stored and may later be recalled. [Anthropic][LangChain] |
| Persistence | The window itself is not cross-session storage. | Retention can span one session or many, depending on the design. [Anthropic][LangChain] |
| Example | Keep a relevant manual excerpt within the current request. | Save a learner preference for a future lesson. [Anthropic][LangChain] |
| Boundary caveats | Larger capacity does not make all stored information visible. | Stored information is not automatically included in the next input. [Anthropic][LangChain] |
Evidence for this distinction
- Context windows ↗ (opens in new tab)Anthropic · Publication date unknown
Relevant section: Understanding the context window - Memory overview ↗ (opens in new tab)LangChain · Publication date unknown
Relevant section: Short-term memory; Long-term memory
Last editorial review: 2026-09-13.