Salience, Ranking, and Metabolism

01Abstract

Long-running agent memory systems commonly treat “importance” as a single quantity. We argue it conflates three distinct signals.

A scoring formula written in 2023 for a simulation demo—relevance plus recency plus importance—moved into production systems unexamined, letting the three jobs of “importance” (sedimentation, ranking, retirement—later separated as salience, ranking, and metabolism) strangle one another inside a single signal.

Our position rests on a service chain: memory serves the LLM; the LLM serves the human—and memory must be memory: self-describing and self-metabolizing, its salience and retirement decided by data and lifecycle rather than one-shot human-delegated verdicts. Skip the middle link and even the most elaborate memory system drifts toward RAG.

Since March 2026, over about seven months of production telemetry on a single conversational system, we argue that “importance” is at least three orthogonal signals that are best kept physically separated. We document three production failures of conflation, contribute a signal taxonomy with four invariants (I0–I3), seven design laws, and two same-day audit tools.

No benchmark supremacy is claimed; all evidence comes from our own failures, of which we have plenty.

02Three production failures

93× → 2.2×

Semantic skew hidden inside one counter, before vs. after splitting the field

0.999 → 0.101

A structure-blind quality metric reading “healthy” while the store degraded

88% → 73%

Raw items that are “islands” (zero converging derivation) — and still declining without cleanup

Gini 0.960

Access-distribution inequality; the top 1% of nodes absorb 51.9% of all accesses

03Figures

Figure 1: three-signal architecture — **Figure 1.** Three-signal architecture: data, service, and lifecycle domains are physically separate, with walls (I1, I2) between them.

Figure 3: positive-feedback loop — **Figure 3.** The positive-feedback loop under conflation, and how the write-back wall (I1) cuts it.

Figure 7: Lorenz curve — **Figure 7.** Lorenz curve of the ranking-layer access distribution (Gini 0.960).

04Two audits you can run today

The Silence Test. Suspend all retrieval for T days (or replay history offline while intercepting writes), then recompute your “importance / quality” metrics. If the numbers do not move, you measured data; if they drift, you measured behavior.

The RAG Test. Fix, in advance, a workload and an outcome criterion; replace your “memory layer” with a good-enough search engine. If behavior shows no material difference on that criterion, the layer was retrieval only—not memory.

05One honest boundary (and a correction)

Silencing retrieval shows zero drift because salience is computed from structural edges only; retrieval counters never enter the formula. Salience and access do correlate weakly (Pearson 0.24) — but that is a shared-structural-cause artifact, not a write path, and invariant I1 constrains the write path, not correlation.

We also corrected an earlier claim: the 0.999 quality metric was structure-blind (a ratio over link types, never counting real structure), not “behavior-contaminated”. The corrected scope-aware metric reads 0.101 on the same store.

06Contents

Introduction: for whom does memory serve — the legacy of RAG
Related work: three heirs and a critical wave
Method: signal definitions, invariants, and a feedback analysis
Case study: production failure evidence (single system, longitudinal)
Design laws
Tools: two audits you can run today
Discussion: limitations, validity threats, and why no benchmark
Conclusion: not sublimation, but regression

Note on evidence. All aggregate statistics are released as anonymized distribution plots in the public preprint record. The raw corpus is never released, for privacy. The Silence-Test protocol is a public artifact, free to reuse and cite.