Summarizing and pruning an AI conversation as it runs so the useful information survives and the noise gets dropped. For agencies, it is the fix for long client threads where the AI slowly starts forgetting the brief.
Also known as compaction, context pruning, conversation summarization
Context compaction is the practice of periodically replacing a long, sprawling AI conversation with a shorter reconstruction of what actually matters in it: the decisions made, the constraints agreed, the facts established. Everything else, including tool output nobody needed, false starts, and repeated corrections, gets discarded.
It exists because of context rot. Models get measurably worse as their context fills, well before they hit the advertised limit, so a million-token window is not a promise of a million tokens of good reasoning. Compaction treats context as a budget to be spent carefully rather than a bucket to be filled.
The failure it prevents is the one agency people complain about most: the AI was great for an hour and then started ignoring things it had already been told.
Long projects are where AI degrades. A campaign build touching strategy, copy, legal notes, and revisions is exactly the shape of work that overruns a context window. Compaction is what lets a single working session cover a whole project instead of restarting every ninety minutes.
It forces a useful discipline. Deciding what survives compaction means deciding what actually matters in a brief. Teams that write good compaction rules usually discover their briefs were carrying a lot of material nobody needed.
It shows up as cost. Every token in context gets paid for on every turn. Compacting a bloated thread is one of the few AI optimizations that improves output quality and reduces spend at the same time.
A team runs a six-week rebrand with an AI assistant sitting in the middle of it, holding the brand strategy, the naming shortlist, the legal screen results, and the client’s feedback from four review rounds. By week three the thread is enormous and the assistant starts suggesting names that were already rejected in round one. The team introduces a compaction step: at the end of each week, the assistant writes a structured summary covering the current positioning, the live shortlist with reasons, the rejected list with reasons, and the outstanding questions, and the next week starts from that summary alone. The thread stops growing, the rejected names stop coming back, and the summary doubles as the status document the account lead was writing by hand anyway.
The automations and agents module of the workshop teaches you how to build AI workflows that compress the busywork without taking the craft out of the studio.