Context And History
The Context setting controls which stored conversation events are included in later model calls. It affects continuity, token usage, latency, and how much old information can influence a response.
Context mode does not change the session ID, memory store, available skills, or system instructions. It also does not disable normal scoped facts; those are a separate model input.
Context Modes
Section titled “Context Modes”| Mode | Current behavior | Best fit |
|---|---|---|
| Full Conversation | Includes all supported conversation-history entries stored for the session. | Short or bounded conversations where full continuity matters. |
| Summarized | Includes the stored running summary, when one exists, plus a recent history window. | Long conversations that need continuity with controlled context size. |
| Last N Messages | Limits context to a recent subset. The configured window is an upper bound in the current runtime. | Tasks where recent activity matters more than older context. |
| None | Includes no stored conversation history. | Requests that should not inherit earlier conversation events. |
History Window is shown for Summarized and Last N Messages. It controls the maximum number of stored history entries in the recent-history portion used by those modes. It does not count conversational turns.
A user request, agent response, clarification, skill call, and recorded skill result can each consume one entry.
Current limitation: In Last N Messages mode, the datastore first narrows history around the latest assistant entry before the history strategy applies the configured window. The model may therefore receive fewer than the configured number of entries. Do not rely on it to provide exactly N messages or exchanges.
What Counts As History
Section titled “What Counts As History”Conversation history can include:
- User requests.
- Agent responses and clarification requests.
- Tool requests and tool results.
Other stored events, including plans, memory operations, usage, warnings, errors, and debug records, are not added to model conversation history.
Images are not retained in the event history. A current request may still contain images when the selected model and integration support them.
Full Conversation
Section titled “Full Conversation”Full Conversation is the easiest mode to understand, but its context grows with the supported history entries stored for the session. Use it when conversations are naturally short or when losing an older detail would be costly.
For unbounded chats, monitor token usage and consider Summarized or Last N Messages.
Summarized
Section titled “Summarized”Summarized combines a stored running summary with a recent history window. The
model is given an agent.summary action and instructed to replace the summary
when it produces a final response. The runtime does not independently generate
a summary when the history reaches a particular size.
The summary therefore depends on the model returning a valid summary action. Before the first summary is stored, only the recent window is supplied. Review summary memory records in Agent Studio when an agent appears to forget, compress, or overemphasize information.
The summary follows the Memory node’s configured scope. With User scope, the same summary can be available to that user in another session. Use Session scope when separate conversations must not share summary context.
Last N Messages
Section titled “Last N Messages”Last N Messages is intended to keep only recent history in model context. In the current runtime, the configured window is a maximum rather than a guarantee because history is narrowed before the configured limit is applied.
Older events remain in the event store for inspection even when they are not sent as conversation context.
Choose a window large enough to preserve the immediate task, tool calls, and follow-up answers.
None disables stored conversation history for model calls. It does not remove all runtime context:
- System instructions are still supplied.
- Available skills are still supplied.
- Normal scoped memory facts are still supplied automatically.
- Runtime action schemas and current date and time are still supplied.
- The current request is still supplied.
Use None for one-shot transformations, classification, extraction, or other requests that should not inherit earlier conversation events. Use a separate Memory configuration or avoid storing facts when requests must not inherit saved information either.
Context And Privacy
Section titled “Context And Privacy”More context can expose more user messages, tool results, and remembered data to the model provider. Choose the smallest context mode that still supports the task.
Normal facts are automatically included in model context regardless of the selected Context mode. Do not store passwords, API keys, authentication tokens, or other secrets as agent facts. Treat debug logs and event exports as sensitive.