Microsoft Patent Filing Describes AI Chatbots That Retain Full Conversation History via Summary

By Patentlyze Team · Updated Jul 3, 2026

AI chatbots have a well-known weakness: they forget things. Microsoft is patenting a fix that keeps a live, size-limited summary of your entire conversation and feeds it back into every new response.

How Microsoft's AI 'whiteboard' keeps your conversation on track

Imagine you're using an AI assistant to plan a project, and thirty messages in, it completely forgets what you decided in message five. That's a real, frustrating limitation of today's AI chat tools. It happens because these models can only read a fixed amount of text at once, and long conversations eventually overflow that limit.

Microsoft's patent describes a system that solves this by automatically creating a "whiteboard" in the background. As you chat, a second AI reads the full conversation and writes a concise, size-capped plain-text summary of everything important. That summary gets attached to each new question you ask, so the model always has the key context, even if the actual chat history is too long to include.

The result: the AI assistant you're talking to behaves as though it remembers the whole conversation, not just the last few exchanges. You get more consistent answers without having to re-explain yourself.

How the whiteboard gets built and fed back to the model

The patent describes a computing system that sits between you and a large language model (an AI that generates text responses). At its core, the system solves the context window problem, which is the hard limit on how much text an AI model can read at once.

Here's the sequence the patent lays out:

• You send a message through a chat interface.
• A separate command goes out to a generative model (which could be the same AI or a different one) asking it to produce a plain-text "whiteboard" based on everything that's happened in the conversation so far.
• That whiteboard has a strict size limit baked into the instruction, so it never grows too large to include in a prompt.
• The system then combines the whiteboard with your new message to build a fresh prompt and sends that to the main AI model.
• The AI responds based on both your current question and the condensed history on the whiteboard.

The whiteboard is described in the patent as "natural language text data" (plain English summaries, not encoded data), which means the AI generating it is essentially writing notes about your conversation in a form another AI can easily read and use.

What this means for long AI conversations in Microsoft 365

For anyone who uses AI assistants for anything longer than a quick lookup, this is a direct quality-of-life fix. Long threads in tools like Microsoft Copilot or a future Teams AI assistant are exactly where context loss hurts most. If the system works as described, you'd be able to have a multi-hour planning session with an AI and never have to repeat yourself because it "forgot" an early decision.

From a strategy standpoint, this also signals that Microsoft is thinking carefully about AI assistants embedded in productivity workflows, where conversations can span hours and dozens of turns. Getting memory right is a prerequisite for those use cases, and this patent is a concrete technical approach to doing that inside the product layer rather than waiting for model providers to expand their context windows.

Editorial commentary on a publicly published patent application. Not legal advice.