Every token admitted to a full context window evicts one already there

The post's split between memory and a cache is a claim about which verbs you get. Memory gives you append, plus the promise that what you put there is still there later ^[7]. A cache gives you admission, eviction, and a miss whose cost depends entirely on which entry left ^[7]. A loop that only implements append has an eviction policy anyway. It is whatever the runtime does when the request exceeds the limit ^[5]. That is inconvenient for the many agent loops written as though there were an append-only mode.

The figures in the post are turn counts. Append is fine at six turns and catastrophic at two hundred, because the window does not grow to match and something the developer did not choose starts falling out ^[8]. On turn one you do not know which of today's facts will matter on turn eighty ^[3]. That is roughly 33 times as many turns competing for the same fixed token budget ^[17].

The one external result the post leans on is "Lost in the Middle" (Liu et al., 2023). It is cited for the finding that retrieval accuracy over a long input degrades measurably when the relevant fact sits in the middle of the context instead of at the start or end ^[9]. The post does not quote a figure from it. Two conditions have to hold before that result says anything about a given run: the packed input has to be long enough to have a middle, and the fact you need has to land in it. A loop that repacks the window each turn and keeps the working set at the edges sits outside the case the study describes.

Contamination is the failure mode the post documents in detail, and the case is narrower than junk accumulating in the window. It is a failed attempt sitting in the transcript as the most recent turn, quietly serving as a template for the next one ^[11]. The post describes a naive retry that resends the full transcript including the model's own wrong answer. On the post's account, that does not hand the model a clean second attempt; it hands the model a context window holding up a plausible-looking answer as the thing to produce again ^[12]

According to the post, plenty of retries are transient failures where the original context was fine and a resend is correct and cheap ^[13]. The fix was classifying why the attempt failed, and for the semantic failures, scrubbing the specific wrong turns while keeping a named statement of what is now ruled out ^[14]. Keeping the literal failed reasoning around because it might still be useful is, in the post's words, "it's not useful, it's an anchor" ^[15].

Adopting that costs two things. Your retry path needs a failure taxonomy before it needs anything else, and the transcript has to be yours to edit turn by turn: deleting turn 41 while keeping turn 42 is not the same operation as clearing the conversation ^[14].

This is one team's field notes, published on dev.to and originally on Loop and Retry ^[16]. The author says both failure modes on either side of the packing decision have broken real runs ^[19], and the available text breaks off while introducing the second ^[18].