Manage conversation context to preserve critical information across long interactions

Summarise older turns as the conversation grows, to stay inside the window.

Summarisation compresses exactly what must stay exact — the amount, the date, the order number — into approximations, and nothing marks the loss.

Keep a persistent case-facts block outside the summarised history, and summarise around it.

Put the detailed results first and the summary at the end, as a conclusion.

Material in the middle of a long input is the least reliably used, so findings buried there are omitted from the answer.

Lead with the key findings, and give the detail explicit section headers so position works for you.

Keep whole tool responses in context so nothing is lost.

A lookup returning forty fields when five matter spends the window on noise, and it compounds with every call.

Trim to the fields the task needs at the point the result enters context.

Send only the recent turns to keep each request small.

The model loses the thread of the conversation and re-asks what was settled, which reads to the customer as not listening.

Pass complete history for coherence, and control size by trimming tool output rather than dropping turns.

Let each agent pass its full working through to the next.

Reasoning chains crowd out the facts, and the downstream agent spends its budget re-reading how a conclusion was reached rather than using it.

Return structured key facts, citations and relevance — not the narrative that produced them.