A support agent in Foundry Agent Service keeps one conversation per customer, and some…

How it works

In Foundry Agent Service, a conversation stores the items of an exchange (messages, tool calls, outputs). When a conversation is used to generate a response, with or without an agent, the full conversation is provided as input to the model.

The runtime components page has a Note for long conversations: if the conversation exceeds the model's supported context size, the model automatically truncates the input context. The conversation itself isn't truncated. Only a subset of it is used to generate that response.

So nothing is deleted or moved. Every item stays stored, and the request still succeeds, but the model may not see the oldest items. If older details matter, that's a design decision for you: summarize, start a fresh conversation with a recap, or use memory deliberately.

Memory (preview) isn't an automatic overflow area. An agent reads and writes the memory store in only two ways: when you attach the memory search tool to a prompt agent, or when you call the memory store APIs directly.

The other options, and when you'd use them

• The service deletes the oldest items: The conversation itself isn't truncated. Only the model's input is.
• Older items move into memory automatically: Memory is used through the memory search tool or the memory store APIs, not as a spill area.
• The request fails: The model truncates the input context automatically and still generates a response.

Exam tip

Conversation too long = input context truncated by the model, stored conversation unchanged.

References

• Build with agents, conversations, and responses
• What is Memory? - Microsoft Foundry