Cheap adoption and weak retirement accumulate cost

Type: kb/types/note.md · Tags: document-system, kb-maintenance

Cheap adoption and weak retirement create a one-sided lifecycle. A contributor can add an index, tag, namespace, template, schema field, or routing convention through one local change. Retiring it later must account for consumers, dependencies, replacements, and migration. When additions keep using the short path while retirement waits on those distributed obligations, dependants accumulate around installed structures faster than obsolete or overlapping structures leave. The imbalance turns initially useful additions into routing, maintenance, and migration cost.

Here, cheap adoption means that the adopter can install a structural choice without first coordinating every later consumer. Weak retirement does not mean that deletion is technically impossible. It means that evidence of obsolescence has no reliable path to a scoped disposition and the downstream changes that disposition requires.

Dependencies amplify the asymmetry

Each dependant can make an adopted structure more useful now and more expensive to remove later. That expense does not establish that the structure transfers to new questions. Current-task fit alone does not warrant costly structural entrenchment separates those propositions: adoption by more tools or notes raises migration cost without earning wider scope.

When a new question no longer fits the installed structure, adding a second structure can still be cheaper for the current author than retiring or reshaping the first. Repetition produces overlapping taxonomies, drifted indexes, competing routes, duplicated mappings, and consumers that disagree about which convention is current. Every added dependant then raises the cost of consolidation. The system can therefore preserve locally rational additions while making later global repair progressively harder.

The cost comes from the lifecycle asymmetry, not from addition by itself. A new structure whose scope is distinct, whose dependants stay bounded, or whose retirement path is already operative need not create the same accumulation pressure.

This mechanism differs from a backlog of defective artifacts. Maintenance capacity must match harmful-artifact inflow concerns retained material that is already harmful or becomes harmful. Here, every installed structure can remain locally sound; the cost can arise only from their combined routing, duplicated maintenance, dependencies, and eventual migration.

Retirement must become an operative path

Retirement needs more than a delete operation. A usable path must connect a signal that a structure has outlived its role to a scoped target, dependency evidence, a decision owner, a disposition, and any required migration. Lineage supplies the source-dependency information needed to invalidate, regenerate, or retire derived artifacts. The broader lifecycle requirement includes redaction, supersession, decay, retirement, and relaxation, but lifecycle metadata becomes management only when a process consumes it.

Named owners, deprecation or supersession state, dependency checks, and scheduled audits can interrupt the asymmetry only when they feed that path. A warning nobody must resolve, a deprecated mark no router reads, or a dependency list with no migration decision leaves retirement weak. Strong retirement need not make removal automatic or as cheap as adoption; it makes recurring evidence capable of reaching a decision and completing its consequences.

Once a committed artifact already has a chosen disposition, Retire an artifact supplies a concrete extraction, inbound-reference, redirect, and freshness-retirement procedure. That procedure starts with a named target. It does not supply the earlier signal, decision owner, or rule that makes accumulated structure enter retirement review.

Bounded observation from agent-memory systems

The comparison of 148 code-grounded agent-memory systems reports 139 automatic writers. Nineteen perform no curation operation, while only seven perform all seven tracked lifecycle operations. This is bounded evidence that complete lifecycle machinery is rare in that reviewed corpus even when writing is automatic. It does not measure the cost caused by missing operations, establish a universal wiki law, or show that every partial lifecycle is inadequate.

Scope

The claim predicts pressure, not inevitable growth. A short-lived system, a stable question set, isolated dependencies, or a retirement path that keeps pace with additions can prevent the cost from accumulating. Overlap can also be deliberate when scopes and routing remain explicit.

The claim does not require every adopted structure to prove permanent value in advance. That would erase the benefit of cheap experimentation. It requires lifecycle capacity to be assessed separately from write capacity: making admission easier without a corresponding way to discover, decide, and execute retirement changes the long-run cost of the system.

The corpus count is corroboration of an asymmetric capability distribution, not evidence that the mechanism caused observed maintenance failures. Target-side evaluation would need histories of additions, dependants, retirement attempts, and resulting routing or migration cost.

Open Questions

  • Which signals show that a structural element has outlived its question, and what rule should make that signal trigger review or retirement rather than passive accumulation? Operational signals that a component is a relaxing candidate develops signals for the narrower case of a brittle proxy component, but not for structural obsolescence in general. Automating KB learning is an open problem places retirement among the judgment-heavy mutations whose oracle remains unresolved.
  • Which observable measures distinguish harmless structural variety from overlap that is already imposing routing, maintenance, or migration cost?

Relevant Notes: