What Copilot Studio uses for memory
Copilot Studio agents carry context in two documented places: their instructions and their knowledge sources. Knowledge sources ground the agent's responses in enterprise data (Microsoft documentation). You can incorporate knowledge at the agent level, in the Knowledge page, or at the topic level with a generative answers node (Microsoft documentation).
AI suggestions for knowledge sources and tools can be accepted or dismissed, and they do not persist beyond the current session (Microsoft documentation). You can change an agent's name, description, and instructions after Copilot Studio finishes provisioning it (Microsoft documentation).
Set the agent's instructions
Start with a description in your own words. Copilot Studio then generates the name, description, and instructions for the agent (Microsoft documentation). The description can have up to 1,024 characters (Microsoft documentation). After Copilot Studio finishes provisioning the agent, you can change the name, description, and instructions (Microsoft documentation).
The documentation also recommends citation instructions. In the agent's instructions, ask the model to always cite its sources (Microsoft documentation).
Add knowledge the agent can reach
Knowledge sources ground the agent's responses in enterprise data (Microsoft documentation). The Documents knowledge source searches documents uploaded to Dataverse and returns results from the document contents (Microsoft documentation). Pick sources that match where your information already lives.
If answers should depend on who is asking, use agent user authentication. With it, when a specific user asks a question, the agent only surfaces content that the specific user can access (Microsoft documentation).
Know the source limits
Each source type has caps that depend on the mode. Public websites support up to 25 websites in generative mode and four public URLs in classic mode. SharePoint supports 25 URLs in generative mode and four URLs per generative answers topic node in classic mode. Dataverse is unlimited in generative mode, and classic mode allows two Dataverse sources with up to 15 tables each (Microsoft documentation).
Two details help at scale. Files uploaded to the agent are not counted in the 25 knowledge source search limit (Microsoft documentation). When there are more than 25 different knowledge sources, generative orchestration filters them using an internal GPT model (Microsoft documentation).
Make grounding strict
The "Allow ungrounded responses" setting, in the Knowledge section of your agent's Generative AI settings, manages whether your agent can respond by using only the model's general knowledge (Microsoft documentation). Turn it off and the agent blocks any response generated in a turn where it did not use a knowledge source or tool (Microsoft documentation).
With the setting off, the agent returns an answer from a knowledge source only when the answer includes an in-text citation to that source (Microsoft documentation). One limit to plan around: citations returned from a knowledge source cannot be used as inputs to other tools or actions (Microsoft documentation).
Connect an MCP knowledge server
Model Context Protocol (MCP) helps you connect to existing knowledge servers and data sources directly within Copilot Studio (Microsoft documentation). You must turn on generative orchestration to use MCP (Microsoft documentation). MCP resources are file-like data that an agent can read for more context, such as API responses or file contents (Microsoft documentation).
KnowMan provides a hosted remote MCP server at https://api.knowman.io/mcp using Streamable HTTP. It signs in with OAuth or accepts an API key sent as an Authorization Bearer header, and it exposes six tools: kb_search, kb_ask, kb_get_item, kb_save, kb_recent, and kb_memories (KnowMan MCP). KnowMan also runs as a local stdio server with npx -y @knowman/mcp on Node 22 or newer (KnowMan MCP).
KnowMan as a personal knowledge base
KnowMan is a personal knowledge base. It saves articles and links, reads and tags them, answers questions with citations, and keeps memories such as claims, decisions, and playbooks. Its hosted MCP server exposes the six tools above, including kb_search for saved items and kb_ask for cited answers (KnowMan MCP).
Copilot Studio documents MCP for connecting to existing knowledge servers and data sources directly, and generative orchestration is required to use it (Microsoft documentation). Separately, KnowMan documents a hosted remote MCP server at https://api.knowman.io/mcp (KnowMan MCP). For the wider pattern, see a knowledge base for AI agents, persistent memory for AI agents, and a personal knowledge base.
Questions people ask
Where does a Copilot Studio agent keep its context?
In its instructions and its knowledge sources. You can change an agent's name, description, and instructions after provisioning (Microsoft documentation), and knowledge sources ground responses in enterprise data (Microsoft documentation).
How many knowledge sources can a Copilot Studio agent use?
It depends on the mode and the source type. Public websites allow up to 25 in generative mode and four in classic mode, while Dataverse is unlimited in generative mode (Microsoft documentation). Files uploaded to the agent are not counted in the 25 knowledge source search limit (Microsoft documentation).
What does Allow ungrounded responses do?
It manages whether your agent can respond by using only the model's general knowledge (Microsoft documentation). When it is off, the agent blocks any response generated in a turn where it did not use a knowledge source or tool (Microsoft documentation). Answers from a knowledge source then need an in-text citation (Microsoft documentation).
Can Copilot Studio connect to an MCP server?
Yes, Model Context Protocol helps you connect to existing knowledge servers and data sources directly within Copilot Studio (Microsoft documentation). You must turn on generative orchestration to use MCP (Microsoft documentation). KnowMan runs a hosted MCP server at https://api.knowman.io/mcp with OAuth or an API key (KnowMan MCP).