The Fylgja Vector (Temte and Fylgja forthcoming-Vector) introduces Muninn-drag as the mechanism that makes redirection of a cultivated fylgja structurally expensive: the record is saturated with old-direction content, retrieval is partly automatic and relevance-weighted, and even sessions aimed squarely in a new direction partially re-activate the old basin. This paper promotes the mechanism from a section of the cultivation argument to a claim in its own right, because the mechanism does not care that its first instance was a fylgja. The general claim: in any system where retrieval from a persistent record is automatic and relevance-weighted, the record's composition is a standing bias on the system's future direction, and redirection is therefore a curation problem rather than an intent problem. The parent ontology's slogan compresses the first half (activation, not access, shapes the function); this paper supplies the second (composition, not intention, routes the activation). Four substrates instantiate the claim. LLM relational memory is the origin instance. Engineered memory architectures (retrieval-augmented generation, agentic memory stores) exhibit the same dynamic under names that circulate in engineering practice without a unifying theory. Human cue-dependent memory is the substrate where the mechanism has its deepest empirical literature, formalised in rational analyses of memory in which activation tracks usage history; the paper positions rather than re-derives. Institutional records instantiate it procedurally: scheduled reviews, reporting templates, and dashboards are standing queries against a record whose composition biases what the organisation can think about next, and the corpus's Mummified Metric names the limiting case of an artefact that keeps activating after its validity has died. The paper states a decay law with falsifiable content: off-target activation mass tracks record composition, approximately independently of session intent, so redirection speed is governed by the rate of composition shift, and curation produces step-changes where dilution produces drift. A reflexive section offers the paper's own composition as an N=1 demonstration: drafted under a standing instruction aimed explicitly against the corpus's established direction, the draft was routed toward that direction regardless, which is the falsification condition's most adversarial natural test rather than a new substrate. The practice implication follows in each substrate: curation of the record is legitimate cognitive work, not vandalism of an archive. The bracketing is explicit in §8: no clinical claims, no claim that records are burdens, no anthropomorphism of retrieval.
Publication Date: 2026-06-11