PRODUCTION SCENARIO
A technical support agent has accumulated 85,000 tokens over 25 turns diagnosing a network outage. IP addresses, firmware versions and verified test results are exact strings scattered through long raw diagnostic dumps the agent has already read. Every verified fact has to be available on every turn, the earliest findings are the settled ones, and latency is already unacceptable, so the team will not add a lookup on the path to a reply.
Which compaction strategy relieves the pressure without losing those facts?
Answering here is anonymous. Nothing is saved unless you sign in.
Show answer and explanation
Answer: A structured record of verified facts, carried forward with the raw output cut
Separating a durable state record from disposable tool output fits every constraint at once: exact values live in a block carried forward verbatim, the dumps that have already been read are pruned, and nothing is fetched at reply time. A narrative summary is right when the thread of reasoning matters more than exact values, and it is how an IP address or a firmware version gets rounded off. A sliding window is right when the task is recency-dominated, and here the oldest findings are the settled ones. A vector store is right when the corpus dwarfs the window, and it returns the top passages rather than all of them while adding the lookup the team has ruled out.