Agent memory rots by accumulation, and the missing primitive is a supersession key

Supersession is the mechanism people will copy from this piece, but the part that decides whether it works is the key. Marking a fact stale requires that a write be reducible to an entity and attribute pair, something like `user.language_preference`, because that pair is what you can look up and what you can point a `supersedes` reference at ^[7]. Embeddings give you no such handle. Two sentences about language preference sit next to each other in vector space, and nothing in the geometry records which one replaced the other ^[5]. That moves the design decision to whoever owns extraction on write, and every fact the extractor fails to key is a fact no policy can govern.

The author's own example is more lopsided than it looks. TypeScript arrives in session one, the Python correction in session forty ^[5]. That leaves up to 39 sessions in which the stale preference could be restated and reinforced, against a single mention of the fact that replaced it ^[1], which is the concrete shape of the observation that frequency and recency are not the same signal and most memory layers track only one ^[6].

The second mechanism is worth reading carefully, because it does not evict anything. Facts unretrieved for N sessions are moved out of the hot index rather than deleted, so they remain recoverable but stop polluting top-k ^[8]. That is an indexing tier, not a bin, and it carries a loop: if salience is a function of how often a fact is retrieved and used downstream ^[8], then a fact that retrieval never surfaces never accrues salience, so the ranking failure that buried it is also what nominates it for archival ^[2]. A supersession pointer at least records an intentional decision. Retrieval-driven salience records the retriever's mistakes and then acts on them.

The cost argument is the strongest thing here and the least measured. Deferring hygiene to read time means over-fetching and paying tokens for the model to reconcile contradictions it is not well suited to reconcile, and that bill grows with the contradiction count until it outpaces the context budget ^[10]. Against that, a targeted lookup on the same entity attribute key at write time is described as orders of magnitude cheaper ^[10]. Neither side of that comparison comes with a number; the extraction pass is called cheap and the saving is asserted, not measured ^[13]. The long-context rebuttal, which the author concedes is incomplete rather than wrong, breaks off mid-sentence in the material supplied to us ^[12], so the case for simply carrying everything in a million tokens is not tested here either.

What survives the missing measurements is the diagnosis. An agent that re-asks a question it already has the answer to, or states a superseded fact with confidence, is not failing at retrieval, and reranking or a larger top_k will not reach the defect ^[3][4]. The store has no representation for a fact going bad ^[4]. That is a schema problem, and it is cheaper to fix before a few hundred sessions of history exist than after.