When you chat with Copilot in Microsoft 365 apps, the assistant may forget what you said a few messages ago. This happens because Copilot has a limited context window that resets after a certain number of turns. The context window is the amount of recent conversation history the model can reference at once. This article explains the technical reasons for context loss, how the context window works, and what you can do to keep Copilot on track.
Key Takeaways: Copilot Context Window Limits
- Copilot context window limit: Copilot in Microsoft 365 retains roughly the last 10 to 15 user messages and responses before older context is dropped.
- Session time-out: Inactive chats reset after 30 minutes of idle time, clearing all prior context.
- Manual context reset: Starting a new chat or clicking the new conversation button immediately clears all previous context.
How the Copilot Context Window Works
Copilot uses a large language model that processes text in chunks called tokens. A token is roughly four characters or three-quarters of a word in English. The model has a maximum token limit for each request. When a conversation exceeds this limit, the oldest tokens are dropped to make room for new ones. In Microsoft 365, Copilot is configured with a context window of approximately 4,000 tokens. That translates to roughly 3,000 words of conversation history. Once you exceed that limit, Copilot no longer sees the earliest messages in the thread.
The context window includes both your messages and Copilot responses. If each turn averages 100 tokens per message, the window holds about 20 turns. After that, the model forgets the first few exchanges. This is not a bug. It is a design constraint of the underlying model architecture. The same limitation exists in ChatGPT, Claude, and other large language models. The key difference is that Microsoft 365 Copilot enforces an additional session time-out. If you stop typing for 30 minutes, the entire session resets and all context is lost.
Another factor is the grounding data Copilot retrieves from Microsoft Graph. When you ask about a file, an email, or a calendar event, Copilot fetches that data and inserts it into the prompt. This data consumes tokens too. A long email chain or a large document can fill the context window quickly, causing earlier parts of the conversation to be dropped.
Steps to Prevent Copilot From Losing Context
You cannot increase the context window size. But you can change how you interact with Copilot to reduce the chance of context loss.
- Keep each message short and focused
Break a complex task into several small requests. For example, instead of asking Copilot to summarize a 50-page document and then draft an email based on that summary in the same message, ask for the summary first. Then in a new message, ask for the email. This keeps each turn lean and preserves context longer. - Repeat key context in every message
If you need Copilot to remember a specific detail, restate it in each message. For example, say: “Based on the Q3 sales report that we discussed, create a bullet list of the top five products.” This ensures the detail stays in the active window even if earlier messages are dropped. - Use the new conversation button for unrelated topics
Each time you switch to a different subject, click the new conversation button. This resets the context intentionally and prevents old data from mixing with new requests. In Copilot for Microsoft 365, the new conversation button is located at the top of the Copilot pane. - Attach relevant files instead of describing them
When you need Copilot to work with a file, use the attach file button in the Copilot pane rather than describing the file in text. Attached files are processed as separate data sources and do not consume conversation tokens in the same way. This leaves more room for chat history. - Monitor the session timer
If you step away from your desk for more than 25 minutes, Copilot may reset the session. To avoid losing context, send a simple message like “continue” before the 30-minute mark. This resets the idle timer and keeps the session alive.
If Copilot Still Loses Context After These Steps
Copilot forgets details after only 3 to 4 messages
If Copilot loses context after only a few messages, you may be sending very long messages. Check the length of each message. If a single message contains more than 300 words, it consumes a large portion of the token budget. Shorten your messages to under 150 words each. Also check if Copilot is pulling in large attachments automatically. Go to Copilot settings and disable automatic data grounding if it is enabled. This prevents large files from eating up the context window.
Copilot suddenly starts a new topic without being asked
This usually happens when the session times out due to inactivity. If you see a greeting message like “Hello, how can I help you today?” in the middle of a conversation, the session has reset. The only way to avoid this is to keep the conversation active by sending a message every 20 minutes. If you need to step away for longer, copy the key details from the chat and paste them into the first message of a new session.
Copilot returns results that ignore earlier instructions
When Copilot ignores formatting or tone instructions you gave earlier, the context window has likely dropped those initial messages. To fix this, include the instruction again in the current message. For example, instead of saying “finish the list,” say “finish the bullet list using the formal tone I asked for earlier.” This restates the instruction within the active window.
| Item | Copilot in Microsoft 365 | Copilot Pro (Chat Mode) |
|---|---|---|
| Context window size | 4,000 tokens (approx 3,000 words) | 8,000 tokens (approx 6,000 words) |
| Session time-out | 30 minutes of inactivity | 60 minutes of inactivity |
| File attachment token cost | Attached files use separate token pool | Attached files use conversation tokens |
| Automatic data grounding | Enabled by default (consumes tokens) | Not available |
| Manual context reset | New conversation button | New chat button |
Conclusion
Copilot loses context between messages because its 4,000-token window drops the oldest content when you exceed that limit. A 30-minute inactivity timer also resets the session entirely. You can reduce context loss by sending shorter messages, repeating key details, and attaching files instead of describing them. For best results, start a new conversation for each separate task and keep your chat sessions under 15 turns.