The Homeostatic Substrate Hypothesis Persistent Agent Memory as Ecological Tissue

Abstract
A recent experimental study of minimally directed embodied agents (Cloos et al., 2026, arXiv:2610.07130)
produced, alongside its intended findings about machine play, an unintended result of equal or greater
significance: agents whose learning substrate is persistent, human-readable text exhibited dynamics —
accumulation, entrenchment, deletion-sensitive function, and self-correction — that the homeostatic-failure
literatures of physiology and ecology would recognize, and that we propose to study under a homeostatic
framing (a framing to be tested, not presumed: per the Round-3 accuracy audit, the source establishes memory-
dependent behaviour, not a homeostatic mechanism). Verified-tagged false beliefs accumulated in memory,
resisted revision, and impaired downstream capability; targeted deletion of a false belief restored function; and
agents showed both entrenchment dynamics (a false conclusion hardening into advice against further
exploration) and self-healing dynamics (public invalidation of a celebrated but impossible result). We propose
the Homeostatic Substrate Hypothesis (HSH): persistent agent memory behaves not merely as a store but as
regulatory tissue — it accumulates allostatic load, entrenches false coherence, resists revision under continued
operation, and heals preferentially by subtraction rather than addition. The hypothesis is stated over a formal
foundation imported from three dated programme documents — Bounded Communication Systems Theory
(channel capacity, homeostatic cost, fold-bifurcation threshold), Bond Sign Management (bond sign, basin
coherence, second-order bonding), and the CDR cycle (coherence–decoherence–recoherence, condensation
nucleus, stuck cycle) — so that “homeostatic” functions as constraint, not metaphor. We formalize a
measurement programme around a Memory Drift Index (MDI) — the exposure-corrected rate of confidence-
tagged memory invalidation, stratified by tag class, with invalidation grounded by pre-registered ablation,
3
consultation, and tag-downgrade criteria so the metric is not circular — and we state explicitly that MDI is
proposed as the agent-memory estimator of the Homeostatic Drift Integral (HDI) defined in Bond Sign
Management v1.2, whose open challenge of domain-level operationalization it is proposed to answer — a
proposal whose validity is itself a pre-registered test (the S4 correspondence arm), not something the present
data establish. Four pre-committed hypotheses with funeral conditions follow; the EH/TH mapping is stated
with its own kill condition; and one engineering consequence is derived: agent memory requires an immune
system (specification: LoopGuard, deposited as SIP Code), not merely a larger store. The gap addressed is
integration, not invention: the components (memory architectures, belief-revision failures, entrenchment
phenomena) exist in the literature; a homeostatic account of their interaction in long-horizon agents does not.