What is context rot — and what does compaction actually lose?

Context rot is the quality slide inside one long session: as the window fills, earlier decisions get crowded out, attention thins, and the agent contradicts its own hour-old work. Compaction trades detail for room — a summary keeps the shape of what happened and quietly drops the constraints, exact values and reasons that made the decisions right.

the mechanics a summary preserves narrative and loses specifics

Every turn, the whole conversation rides along as context. Long sessions approach the window's limit, so harnesses compact: older turns are replaced by a model-written summary. The summary says "we set up auth" and loses which flag, which version, and why the obvious alternative was rejected. The agent then re-derives specifics — sometimes differently than it decided two hours ago.

why a bigger window does not end it present in the window ≠ applied to the work

Longer windows delay compaction; they do not remove the decay. Attention over a very long context is not uniform — material in the middle competes to be noticed, and a decision being present is not the same as a decision being applied. The failure is architectural: the window is working memory, and working memory is the wrong home for durable truth.

what belongs outside the window then compaction can lose the transcript freely

Durable decisions belong in a record that does not shrink when the session grows: each one atomic, carrying its reason and its source, served back when relevant rather than carried always. Measured against a live project's rule set, recall from the pinned record covered 1.000 of the load-bearing rules; similarity search over raw session text covered 0.094 of the same set — the transcript was never the right place to keep the law.

npx memcell connect