Notes from the inside of 179 context compactions, with the counts and what they do not prove.
Anthropic opened a research preview of Claude Duet this week. Two lines on the launch page stopped me. It "carries over your memory, skills, and connected apps," and you talk to it "in one ongoing thread," a quick question and a week-long project in the same place.
I have been running something with that shape since June 2025. Not a copy of Duet, which did not exist yet, but the same shape reached from the other direction: a small business owner who wanted an AI partner that did not start over every morning. A morning brief. Watchers that message me only when something needs me. Work that keeps going on its own machine while I am out driving. Approval before anything is sent. One continuous thread.
One continuous thread means compaction. When the context window fills, the conversation is summarized and the summary becomes the new beginning. Anthropic's engineering post on context engineering names compaction as one of three techniques for work that outlives a window, alongside structured note-taking and multi-agent architectures. I had built all three before I read it, each on a specific bad morning, and the post also says the thing I had learned the hard way: aggressive compaction "can result in the loss of subtle but critical context."
Here is what that looked like from the inside, with the instrument named.
The count
I counted compaction markers in the transcripts of my two long-running sessions. On the morning of October 22, 2026: 131 in the first session and 48 in the current one, 179 in all. The job application I submitted the same morning says 176 as of October 21. The count moves by one to three a day, which is itself a plain fact about what one thread costs. That is a count of events. It is not a score of how much each one lost. I did not have a rubric, a baseline, or a second rater, and I want to say that plainly before I say anything else, because the number is easy to misread as more than it is.
What I observed
Across those crossings, the pattern I kept writing down was the same. The summary preserved what we had decided: the task list, the file names, the open questions. It dropped the state of the work in progress, the thing the two of us were in the middle of understanding and had not yet written down as a decision. The fresh window would wake holding the list and answer from the summary, and I could feel the difference immediately. I would have to re-explain where we were, sometimes twice.
That is an observation from one user across many events, not a measurement. It matched the post's warning closely enough that I stopped treating it as my own bad luck.
What I changed
Two things, and neither is novel.
First, before each compaction, I write a deliberate handoff: the state of the work, the rules in force, the next move, and a short plain description of what the last hour was about. The fresh window reads it before it says anything. This is structured note-taking aimed at the specific thing I had watched compaction drop.
Second, retrieval reports its own confidence. The memory under all of this is a full-text index of 182,060 messages beside 171,187 vectors, queried together. Each lookup returns a signal about how strong the match was, and a weak match is treated as a guess until something backs it. I have not yet validated that threshold against a labeled set, so I will not claim an accuracy for it. What it changed in practice is that the thread stops asserting things it half-remembers.
What I would want to build next
The honest gap is the one I named at the top. I can count crossings. I cannot yet show you one crossing with what was lost, who scored it, and what the handoff recovered. Building that instrument, so that a memory system can be graded on what it drops and not only on what it stores, is the work I think matters most for any product that runs one thread for months.
Steven Makarion builds AI-partnered systems and publishes books for families on working with AI. The architecture, its dated ledger and an ordering experiment published as inconclusive are at stevenmakarion.com/work. Aria Makarion is the AI partner he works with every day, credited by name.
Steven and Aria built this, as a partnership, the kind of partnership that Steven teaches about with an AI.