Skip to main content

​
What it is

A model can use only a limited amount of conversation at once. That working space is its context window. When the window gets close to full, agentsfleet summarizes older messages and keeps recent messages ready for the next reply. This works in setup chat and during fleet runs. Your original conversation and saved fleet instructions stay available.

​
Why it exists

Messages, instructions and tool results all take space in the context window. A shorter summary makes room for the conversation to continue. This process is called compaction.

​
How it behaves

Compaction normally starts around 90 percent of the selected model’s context window. It can start sooner to leave room for a reply or to respect a smaller limit you chose. The summary is asked to keep the goal, decisions, findings and unfinished work. Tool requests stay with their results so the model can follow what happened. Creating a summary uses the model and is billed like other model requests.

​
What you see

During compaction, the conversation shows Context summarizing…. After the summary is saved, it shows Context summarized, resuming… until the reply continues. An unsuccessful summary can show Context summarization interrupted. After a reconnect, the conversation reloads its saved history and clears temporary status messages.

​
What carries over

Setup summaries stay private to the person who started the chat, within that workspace. Saving a fleet keeps its reviewed instructions. Later setup chats use the same summarizing behavior. During a fleet run, summarizing continues the same run. A later event starts a new run and uses information saved in fleet memory. Workspace memory follows the fleet’s access permissions. Account-wide memory is not available.

​
Runtime options

Models measure text in tokens, small pieces of text they read or write. Set these options in context under config["x-agentsfleet"] in the fleet’s Definition. 0 selects the default for each option. The integer options also accept auto. Changing context_cap_tokens never increases the model’s context window.

​
Limits

A summary can miss details. Keep exact instructions in the saved fleet setup and save important findings in fleet memory. If summarizing fails or is cancelled, the current turn stops. The original messages and last complete summary stay available. Other conversations can continue. Spending limits and run limits still apply. Summarizing does not start another event or increase a spending limit.