Spare capacity, narrow surface: the exposure record of a production agent-memory system

The field's canonical survey of agent memory (218 papers) catalogues evaluation on fixed sets of queries and tasks, with retrieval, memory-quality, response-quality and end-to-end task metrics; its metric taxonomy does not include a census of distinct memory items delivered under production traffic. Retrieval is also conditional on a query having been issued, so an item that no query reaches has an undefined nDCG, not a low one. For 12 weeks we instrumented the two surfaces through which a memory system in operation delivers content to a fleet of 6 agents: a proactive 10-item brief and on-demand search.

The brief logged 583,763 selected slots, 8.7 times the size of the corpus, enough to serve each of the 67,187 chunks eight times. They hold 1,635 distinct live chunks: 2.43% (1,787 counting the 152 that were served and later deleted; see §4.1). These are logged selections: the log is written before text rendering, and delivery over the whole window was not verified (§3.1). The aggregate capacity therefore did not force this result. It does not follow that the ordering is wrong, only that the number was not produced by a lack of space. The per-session capacity (10 items) is not tested here. (The 99.98% expected coverage under uniform random service, against a 100% maximum, illustrates what the capacity would allow; it is not a recommended policy.) Adding search, 56,288 live chunks (83.78%) have neither a brief-log record nor a positive search counter. This is a lower bound on non-delivery by the brief and by tracked search; non-delivery across all agent-facing search is not established (§3.1). The capacity figure and the 83.78% sit in different universes on purpose: the slack is the brief's, the 83.78% counts the union. Of the 10,899 live chunks with a record on either surface, 9,755 have a positive search counter. That counter records that a chunk was a top candidate of some tracked search call, whoever initiated the call (automated callers included) and whether or not the chunk was then returned: it bounds from above what tracked search returned (§3.1). The brief is the surface that proactively selects content for delivery. Every claim in this paper about mechanism concerns the brief; the 83.78% describes the state and does not assign a cause.

The two channels of the surface freeze, for opposite reasons and neither tied to capacity. The 8 slots of the main pool are ordered by a score whose terms, with one exception, do not decay: the access component is monotone in a counter that only goes up. The 3 chunks present in 100% of the 4,632 briefs of the week rank 1, 3 and 4 by salience among the 149 served chunks that still exist (last accessed 42, 90 and 30 days before the close of the window, 2026-08-28; §4.3.2). With the access term zeroed they fall to ranks 44–46: tracked search traffic from months ago (§3.1) is necessary for their observed salience ranks in the measured counterfactual. The high-pain pin selects them in phase 0, before the quota pass, but it protects only items that the score has already placed there (brief.ts:819-821), and lifting it removes only one of the three from the main set (§4.3.2). The top-10 takes 47.16% of the slots. The other 2 slots are a coverage channel, whose declared purpose is to serve the never-served, and it freezes for a different reason: its eligible population is 108 chunks in a corpus of 67,187 (0.16%, carved out by two path patterns together with the channel's importance floor and 30-day window), and it exhausts that population entirely on every measured day, with 12.4 slots per candidate on a closed day (an unpreserved reading on a partial day gave 5.6, and the two are not comparable; see §4.3.1). On every measured day there was no never-served item left to serve. These coverage numbers were measured on 2026-08-26 to 08-29, inside a period of defective session ingestion (§4.3.1), and describe the channel in that regime, not its steady state. The main pool would respond to a score adjustment, and nobody adjusts it. The coverage channel, designed to compensate for the main pool, has a daily reach bounded by its eligible population (path patterns, the importance/pain floor and age windows), which it exhausted in that regime; within a brief it responds to score only within last_served ties (17/350 here) and then saturates.

The mechanism of the coverage channel is deducible from the code. It orders by a lexicographic comparator (last_served ASC, salience DESC): the score is the subordinate coordinate and decides only within ties of the dominant one. That predicts a ceiling, not a proportional response, for any additive bonus on the score of that channel. We tested with an increasing dose by counterfactual replay over 350 of 350 real brief states, faithful to the serving pipeline. The states are from production, and the intervention was not served: the mode is shadow, so the treated composition is computed and recorded, and what the agent received during the measured window was always the control (§7). Result: a monotone response in each state, saturation at w ∈ (4.0, 4.4], and a ceiling of 4.86% of briefs (measured on the 2026-08-26 corpus, in the same regime, with an empty never-served stratum). The ceiling is also not a constant of the mechanism: under the same rule with another draw of designated items it reaches 7.43% (the draw in force sits at the minimum of the distribution, tied with another), and truncating the timestamp resolution from second to minute or hour takes it to 36% and 80%, without changing a single line of code. The reach of the mechanism is fixed by decisions that nobody took as policy. A third axis: excluding the 25 rows of our own health probes from the serve-state lowers the ceiling to 13/350 (3.71%), with one sensitive state in common between the two arms (§5.7.2).

What we do not claim: no effect on agent behavior, since no downstream outcome is instrumented (§4.5). Nor do we claim that the concentration is wrong. A policy that serves 10 items per session should concentrate; uniform serving is used only as a capacity reference, and its effect on agent utility was not measured; the finding is that the non-exposure is a result of policy and not a capacity limit, hence revisable by a design decision. Collection size, which correlates with exposure (§4.2), may be a proxy for how the type is produced: curation is not ruled out as a common cause. And we claim nothing about the field: this is one system. The generalization of the mechanism is deductive and holds for any ranker with lexicographic order and a bonus on the subordinate coordinate that serves a prefix of that order. How many systems have this shape is an open question, and the executable diagnostic we publish exists so that others can answer it one at a time.